Introduction

Due to human interventions and the use of fossil fuels, the greenhouse gas concentration in the atmosphere has increased and resulted in a global energy disparity (Wentz et al. 2007; Chu et al. 2010; Huang et al. 2011). This in turn has led to what is called climate changes. Change in the climate means a state within a certain period where the statistical distribution of weather patterns is changeable due to the radiative forcing (RF) variations (Al-Mukhtar 2018; Moussa et al. 2018). The sources of these variations arise from the natural, anthropogenic, and environmental process, which influence on the earth’s energy budget (Khadka and Pathak 2016). The RFs are embedding to the climate models as boundary conditions to simulate future scenarios. The total anthropogenic RF was estimated according to IPCC Fifth Assessment Report (AR5) in 2011 as “43 % higher than that reported in the fourth Assessment Report (AR4) published in 2007” (IPCC 2013).

Iraq was characterized as an abundant water resources country among the other regional countries with an annual allocation per capita of 6029 m3 in 1995; however, a significant drop was recorded in the annual allocation to become 2100 m3 in 2015 (Nimah 2008). Restoration project for the marshes in the south of Iraq and the increasing demand on water in the neighboring countries will result in desiccation the Tigris and Euphrates Rivers in 2040 (Al-Ansari et al. 2012). This would in turn lead to some ecological and environmental consequences not only in Iraq but also on the Arabian/Persian Gulf and Kuwait (Al-Ghadban et al. 1999, 2008). Some suggested solutions to overcome the water shortage crises, but not only, are to use non-conventional water resources management techniques (e.g., water harvesting) (Al-Ansari et al. 2012; Zakaria et al. 2013) and by attending a mitigation and adaptation of climate change.

In order to adopt such strategies and to avoid ill-informed decisions about the prospective water resources availability, planners need knowledge about the impacts of climate changes on meteorological parameters at local scale. The global climate models (GCMs) are the main informative tools used for present and future climate change scenarios. However, the resolution of the GCMs is very coarse (usually more than 40,000 km2), and thus, they are unable to accurately model the sub-grid scale climate features of the local-scale areas (Wilby and Dawson 2007). Thus, there is a need for spatial and temporal downscaling from the GCM into a finer scale.

Downscaling technique is defined as the tool to create a link between a regional/local and GCM scales (Wetterhall et al. 2009). The range of local and regional scales is 0–50 km and 50 × 50 km, respectively (Xu 1999). Downscaling can be grouped into two categories: dynamic and statistical downscaling (Christensen et al. 1997; Fowler et al. 2007). Dynamical downscaling (DD) is a process-based technique focused on driving nested high-resolution circulation model called regional climate model (RCM). This model takes the boundary conditions from the coarse resolution data of GCMs and provides information on fine temporal and spatial resolution that able to simulate the physical processes (Giorgi 1990; Jones et al. 1995). On the contrary, statistical downscaling (SD) methods associate local-scale with large-scale variables through empirical/ statistical relationship by using multiple linear regression and the stochastic weather generator (Gebremeskel et al. 2005; Diaz-Nieto and Wilby 2005; Gagnon et al. 2005; Wilby et al. 2006). Although these methods are characterized as faster, and computationally inexpensive (Wilby et al. 2000), they have shortcomings represented by assuming a constancy of the statistical association in the future.

Weather generators are stochastic models employed in SD methods to synthetically generate a long series data, reproduce missing data, and synthesize different realizations of the same data through a random number generator (Wilks and Wilby 1999). Statistical downscaling model (SDSM) is being widely used throughout the world as a statistical downscaling tool (Huang et al. 2011) to downscale most of climate variables (such as temperature, precipitation, and evaporation) which could be used in further assessing of hydrologic responses.

The objectives of this study were to investigate the ability of statistical downscaling model in reproducing the meteorological parameters (i.e., precipitation and temperature) and, hence, to analyze the impact of future climate changes (2011–2100) on precipitation and temperatures in several station of Iraq. To this end, the SDSM was calibrated and validated against daily observations from 12 stations scattered from the north to the south using NCEP/NCAR parameters. Then, the calibrated model was applied to project future precipitation and temperature over three time slices using the Canadian GCM. The outcomes from this study would help decision makers and researchers in better planning for the water resources in Iraq and help in finding ways and means to reduce the effects of climate changes on the inhabitants and the environment.

Study area and data description

Iraq occupies a total area of 437,072 km2. Landforms constitute 432,162 km2, while water forms 4910 km2 of the total area. Turkey borders Iraq from the north, Iran from the east, Syria and Jordan from the west, and Saudi Arabia and Kuwait from the south. The total population in Iraq in 2017 is about 37,139,519. Iraq is composed of 18 governorates. Twelve sites were selected across Iraq to represent as much as possible major climatic regions in the country (Fig. 1). As such, the selected stations extend from the north to the south of the country where most of the agricultural and urban areas are present. The investigated stations are Baghdad, Kirkuk, Mosul, Sulaymaniyah, Najaf, Nasiriyah, Al-Hay, Basra, Zakho, Erbil, Salah ad Din, and Khanaqin. Table 1 presents a detail description about the locations of the investigated stations.

Fig. 1
figure 1

Location of the meteorological stations

Full size image
Table 1 Location details of the 12 stations in Iraq
Full size table

The observed data of those stations were obtained from the Iraqi National Meteorological Organization and Seismology. The data availability is listed in Table 2 with their corresponding mean annual precipitation and temperature. The observed daily rainfall and temperature data of the stations Kirkuk, Mosul, Baghdad, Nasiriyah, Najaf, Al-Hay, Basra, and Khanaqin were available for the period 1961–1990. While for the remainder stations (Sulaymaniyah, Zakho, Erbil, and Salah ad Din), the data were from 1971 to 2000. For better understanding of the climate behavior in those stations, the mean monthly precipitation and temperature were depicted as shown in Fig. 2. As it can be seen from the figure, the mean monthly precipitation of the 12 stations over the respective period ranged from 700 to 50 mm in January to zero in the summer months (Jun, July, August) and September. On the other side, the mean monthly temperature fluctuated from 45 to 30 °C in summer to 20–5 °C in winter. The climate is classified as arid in south to semi-arid in the north.

Table 2 Description of data availability for the 12 stations used in the study
Full size table
Fig. 2
figure 2

Monthly a temperature and b precipitation at weather stations for the baseline period (1961–1990 and 1971–2000)

Full size image

The 26 predictors of NCEP/NCAR reanalyzed predictors with grid resolution 2.5° × 2.5° were freely downloaded from (https://www.esrl.noaa.gov/psd/data/reanalysis/reanalysis.shtml). These predictors represent the observed coarse resolution parameters of past to present time.

The CMIP5-ESMs climate model of the Canadian model for a very high emission (RCP8.5), an intermediate emission (RCP4.5), and very low emission (RCP2.6) scenarios with grid resolution 2.8125° × 2.8125° were obtained from the Canadian Centre for Climate Modeling and Analysis (http://climate-scenarios.canada.ca/?page=pred-canesm2) for the periods of 1961–2001 and 1961–2099. These predictors are available in zip file format and encompass five files inside: CanESM2_historical_1961–2005, CanESM2_rcp2.6_2006–2100, CanESM2_rcp4.5_2006–2100, CanESM2_rcp8.5_2006–2100, and NCEP-NCAR_1961–2005. Due to the presence of inconsistency between the GCM and NCEP resolutions, the NCEP data were interpolated to be consistent with the data grid resolution of CanESM2. The above variables with those from observations were used for calibration, validation, and projection in SDSM.

Methodology

The SDSM

The statistical downscaling was carried out by SDSM. The model was downloaded from http://www.sdsm.org.uk. Wilby et al. (2002) developed the model as a hybrid tool of both the multiple linear regression (MLR) and the stochastic weather generator (SWG). The rule of MLR is to establish a statistical/empirical relationship between large-scale and local-scale variables, and make some regression parameters from the present data. In order to establish a better association with the observed time series, the calibrated parameters from MLR along with NCEP and GCM predictors are fed into SWG to simulate up to 100 daily time series (Wilby et al. 2002; Liu et al. 2009).

The selection of the most suitable predictors among the atmospheric predictors is done through a combination of statistical tools, i.e., correlation matrix, partial correlation, P value, histograms, and scatter plots. The calibrated parameters are optimized through either ordinary least squares (OLS) or dual simplex (DS) (Huang et al. 2011). The model determines the relationship between the predictands and predictors through three kinds of sub-models—monthly, seasonal, and annual. Annual sub-model derives one-regression parameters for 12 months. While, the monthly sub-model represents 12 regression equations of the 12 months. Moreover, conditional and unconditional processes are applied within the model for precipitation and temperature, respectively (Wilby et al. 2002; Chu et al. 2010), where the conditional process depends on an intermediate variable to link the predictors with predictands, while the unconditional process considers the direct link between them (Khan et al. 2006). The methodology followed in this study for downscaling and scenario generation is shown in Fig. 3.

Fig. 3
figure 3

Flow chart showing steps involved in downscaling and scenario generation (after Wilby and Dawson 2007)

Full size image

Screening of probable predictors

One of the most important steps in statistical downscaling is to identify properly the most effective predictors on predictands in what is called screening of predictors (Hewitson and Crane 1996). This process is employed to select appropriate sets of observed predictors from the suite of NCEP/NCAR reanalysis datasets based on several statistical metrics. The choice of predictors differs based on geographical regions considering the properties of the predictor and the predictand (Anandhi et al. 2009). Therefore, in this study, the scatter plots, value of explained variance, correlation, partial correlation (r), and P value approaches were used for obtaining the most suitable sets of predictors from the suite of NCEP/NCAR data at an individual station. As pointed by Wilby et al. (2002), “the explained variance explains the level to which daily variations in the predictand are determined by predictors.” While, the correlation statistics and P values are a measure of the strength of relationship between the predictor and predictand. The smaller P value (P < 0.05) implies a better association between variables.

Calibration and validation

Quality control check function was used to identify whether the data have missing or outlier values. The observed data in each station (predictands) with their NCEP/NCAR were used to calibrate and validate the model. Data was partitioned into two parts: 70% for calibration and 30% for validation. So, data from 1961 to 1981 and 1971 to 1991 was used for the calibration, where the model was calibrated with the screened NCEP predictors using the annual sub-models, for the respective predictands. The unconditional sub-model without transformation was used for temperature and conditional one with fourth root transformation for precipitation. The OLS was adopted as an optimization method. Subsequently, daily dataset for the period of 1982–1990 and 1992–2000 was selected for the validation of each predictand. Validation was carried out using calibration output, and observed NCEP reanalyzed atmospheric variables. Then, SWG was employed to generate 20 ensembles of synthetic weather series. The average of these ensembles was compared with independent observations from the validation period.

Downscaling of future predictands

In this study, the calibrated model was used in further analysis, i.e., assessing future climate changes on temperature and precipitation over three periods by downscaling the RCP scenario predictors obtained from the CanESM2 model. To achieve that, the built-in SWG was employed to generate 20 ensembles of future predictands. Ultimately, the downscaled future climate change predictands in the 2020s (2011–2040), 2050s (2041–2070), and 2080s (2071–2099) were compared with those in the baseline period (1961–1990). The anomaly of monthly precipitation/temperature was obtained from the percentage change/absolute difference of average 20-ensemble future predictand with respect to monthly average of the baseline period. The positive or negative anomaly indicates that there would be increase or decrease of the variable in future periods, respectively.

Evaluation criteria

In order to evaluate the SDSM performance with respect to the observed Tmax, Tmin, and precipitation data, the following three statistical model performance evaluations were applied. These statistical performance measures are mostly used to check the goodness of fit between the observed and modeled (Al-Mukhtar 2016).

Coefficient of determination (R 2)

The determination coefficient or coefficient of determination (Eq. 1) is a measure used to determine the variability in observed data that the model could capture it (Krause et al. 2005).

$$ {R}^2=\frac{{\left(\sum \left[{X}_i-{X}_{av}\right]\left[{Y}_i-{Y}_{av}\right]\right)}^2}{\sum {\left({X}_i-{X}_{av}\right)}^2\sum {\left({Y}_i-{Y}_{av}\right)}^2} $$
(1)

where Xi is measured value, Xav is average measured value, Yi is simulated value, and Yav is average simulated value.

Nash–Sutcliffe coefficient

The Nash–Sutcliffe coefficient (NSE) is a dimensionless model evaluation statistic where the relative magnitude of the residual variance is determined in comparison to the observed variance (Moriasi et al. 2007) and is defined as the following (Nash and Sutcliff 1970):

$$ E=1-\frac{\sum_{i=1}^n{\left({X}_i-{Y}_i\right)}^2}{\sum_{i=1}^n{\left({X}_i-{Y}_{av}\right)}^2}. $$
(2)

Root mean square error

The root mean square error (RMSE) is an error index type of model evaluation statistics (dimensional). The closer value to zero, the better model performance (Singh et al. 2004).

$$ \mathrm{RMSE}=\sqrt{\frac{\sum_{i=1}^n{\left({X}_i-{Y}_i\right)}^2}{n}} $$
(3)

where n is the number of values.

Results

Screening of predictors

The best-correlated predictor variables were selected based on P value and partial r for each station’s predictands. Results revealed that different atmospheric variables affect different local variables. The driving parameters on Tmax and Tmin with significance level < 0.05 were common. In other words, ncepp500-gl (500-hpa geopotential height), nceps850-gl (specific humidity at 850 hpa), ncepshumgl (surface-specific humidity), and nceptempgl (mean temperature at 2 m) were the most effective parameters on temperature at all stations. These parameters seem physically rational because they are highly related to the changes of temperature characteristics through the thermal advection term (Romanowicz et al. 2016). While, the driving parameters on precipitation were varied across the stations (Table 3). For example, in Baghdad station, the most significant parameters were ncepmslpgl (mean sea-level pressure), ncepp1-fgl (1-hPa air flow strength), ncepp500-gl (500-hPa geopotential height), nceps500-gl (specific humidity at 500 hPa), and nceptempgl (mean temperature at 2 m). Whereby, in Basra south of Iraq, in addition to the above parameters, the most significant parameters were ncepp1-ugl (1-hPa zonal velocity), ncepp1-zhgl (1-hPa divergence), and ncepp850-gl (850-hPa geopotential height). The reason behind such as variation might be attributed due to the precipitation heterogeneity (Wilby et al. 2002). The most effective parameters on precipitation in most of stations were the specific and relative humidity. These parameters are associated and highly correlated to precipitation occurrence because their synchronous variation is dependent on the saturated phase of water vapor in the air (Hessami et al. 2008). The significant deriving parameters were subsequently used in the SDSM calibration.

Table 3 Significant predictors from CanESM2 model for precipitation
Full size table

Calibration and validation

Based on station data availability, the observed data was either from 1961 to 1990 or 1971–2000. This data was portioned into two parts: for calibration and validation. Data from 1961 to 1981 and 1971 to 1991 was used for the calibration, whereas data from 1982 to 1990 and 1992–2000 was used for the validation of the model. The statistical performance of calibration and validation results for the 12 spatial stations between the observed Tmax and those counterparts from the downscaling model are listed in Tables 4 and 5, respectively.

Table 4 Statistical performance of Tmax modeling during calibration
Full size table
Table 5 Statistical performance of Tmax modeling during validation
Full size table

It can be noticed from Table 4 that the RMSE, R2, and NSE were ≤ 5.57 °C, ≥ 0.80, and ≥ 0.70, respectively, across all the stations. According to the RMSE values, Kirkuk station has the lowest value (4.65 °C) among the other stations. In overall, results proved a well performance of SDSM in modeling maximum temperature during the calibration period. Thus, those results can be used in further analysis, i.e., validation phase. During the validation period, the statistical performance was almost akin to that in the calibration. The values of RMSE, R2, and NSE were ≤ 6.67 °C, ≥ 0.78, and ≥ 0.67 across all the stations (Table 5) with the lowest RMSE in Kirkuk station as well. Therefore, it can be concluded that the performance of the calibrated model was representative for the entire data set and, thus, can be used in future projection.

Tables 6 and 7 list the statistical evaluation performances of Tmin predictand during the calibration and validation, respectively. It can be noticed from Table 6 that the RMSE, R2, and NSE were ≤ 4.56 °C, ≥ 0.83, and ≥ 0.69, respectively, across all the stations during the calibration. According to this evaluation, it can be judged that the observed and the modeled data were consistent. In other words, the SDSM were sufficiently capable to reproduce the observed Tmin data. During the validation period, the values of RMSE, R2, and NSE were ≤ 5.25 °C, ≥ 0.76, and ≥ 0.69 across all the stations, respectively (Table 7). Thus, the calibrated model could be used for future projections with respect to minimum temperature.

Table 6 Statistical performance of Tmin modeling during calibration
Full size table
Table 7 Statistical performance of Tmin modeling during validation
Full size table

With respect to precipitation modeling, the performance evaluation shows that the observed and the modeled data were close to each other. The values of the RMSE, R2, and NSE were ≤ 0.95 mm, ≥ 0.51, and ≥ 0.60, respectively, across all the stations during the calibration period (Table 8). Therefore, it can be concluded that the model’s performance was satisfactory in presenting a rather well correlation between both data sets and subsequently was used in the validation phase. The values of statistical performance evaluation during the validation are listed in Table 9. RMSE, R2, and NSE were ≤ 1.01 mm, ≥ 0.55, and ≥ 0.60, respectively, across all the stations. Thus, the SDSM was efficiently capable to reproduce the observed precipitation data in such as arid and semi-arid areas.

Table 8 Statistical performance of precipitation modeling during calibration
Full size table
Table 9 Statistical performance of precipitation modeling during validation
Full size table

In general, the downscaled temperature and precipitation values were closely consistent with the observed. Many previous studies (Hassan et al. 2014; Yang et al. 2012, and many others) have demonstrated that the downscaling can better reproduce temperature series than precipitation. However, we conclude that the SDSM is powerful in reproducing predictand values and, therefore, the model efficiently downscaled the maximum-minimum temperature and precipitation during the calibration and validation periods.

Downscaling future climate scenarios

The SDSM model developed for each site was used to predict future daily precipitation and temperature in the sites for the periods of 2011–2040, 2041–2070, and 2071–2100 (hereinafter first, second, and third periods, respectively) based on the RCP2.6, RCP4.5, and RCP8.5 scenarios generated from CanESM2. Side-by-side anomaly results for all the meteorological stations of the three different scenarios are plotted as shown in Figs. 4, 5, and 6 for Tmax, Tmin, and precipitation, respectively, projected by the SDSM model under the three future periods. Tables 10, 11, and 12 tabulate the anomaly values of Tmax, Tmin, and precipitation, respectively, projected by the SDSM model under the three future periods. As it can be noticed, the maximum temperature was projected to increase across all the 12 stations with a minimum increase of 0.04 °C under RCP2.6 in Basra during the first period and maximum increase of 11.3 °C under RCP8.5 in Sulaymaniyah during the third period (Table 10 and Fig. 4). The minimum temperature was also projected to increase in all stations except Basra where the Tmin was projected to decrease. The minimum change in Tmin was 1 °C in Nasiriyah during the first period and maximum increase of 6.7 °C under RCP8.5 in Sulaymaniyah (Table 11 and Fig. 5). On the other side, precipitation was projected to decrease across all the evaluated stations. The maximum decrease was − 68% under RCP8.5 in Sulaymaniyah during the third period and the minimum decrease was − 1% (Table 12 and Fig. 6).

Fig. 4
figure 4

Change in average annual maximum temperature in the future under CanESM2-RCP2.6, CanESM2-RCP4.5, and CanESM2-RCP8.5 scenarios

Full size image
Fig. 5
figure 5

Change in average annual minimum temperature in the future under CanESM2-RCP2.6, CanESM2-RCP4.5, and CanESM2-RCP8.5 scenarios

Full size image
Fig. 6
figure 6

Change in average annual rainfall in the future under CanESM2-RCP2.6, CanESM2-RCP4.5, and CanESM2-RCP8.5 scenarios

Full size image
Table 10 Anomaly of maximum temperature (°C) from RCP scenarios over first, second, and third periods relative to baseline
Full size table
Table 11 Anomaly of minimum temperature (°C) from RCP scenarios over first, second, and third periods relative to baseline
Full size table
Table 12 Anomaly of precipitation (%) from RCP scenarios over first, second, and third periods relative to baseline
Full size table

Discussion

The statistical downscaling model shows a superior performance in modeling daily precipitation and temperature across the 12 studied stations in Iraq with a slight difference in performance among stations. Hessami et al. (2008) argued that “the agreement of simulations with observations depends on the GCMs atmospheric variables used as “predictors” in the regression-based approach, and the performance of the statistical downscaling model varies for different stations and season.” The downscaling procedure of precipitation and temperature of this study was in agreement with other studies (e.g., Wilby et al. 1998; Huth 2002; Gachon et al. 2005), where they proved that using circulation variables (exemplified by geopotential, vorticity, or the wind component) and temperature (exemplified by geopotential heights at various levels and specific/relative humidity near the mid-troposphere and specific/relative humidity) are superior in establishing a satisfactory relationship to that of any single predictor when downscaling temperature and/or precipitation. Presumably, the SDSM would receive more attention in future studies inside Iraq among the other downscaling methods because of “less preprocessing requirements and computational costs” (Tavakol-Davani et al. 2013) in addition to the simplicity in implementation.

The outcomes from this study pointed an increase in temperatures and decrease in precipitation across the country. The greatest increase in annual maximum temperature was detected in Sulaymaniyah station in the north of Iraq at the end of the century 2080s under RCP8.5. While the lowest increase in maximum temperature was in Basrah station at 2020s. On the other hand, the greatest increase in minimum temperature was observed in Salah ad Din station (186 km to the north of Baghdad) at 2080s under RCP8.5. The lowest increase in minimum temperature was found in Nasiriyah (station in the south of Baghdad) at 2020s under RCP2.5. Surprisingly, Basrah station recorded a decrease in the minimum temperature during the three future periods except a slight increase at 2080s under RCP8.5. This could be attributed to the fact that warm and cold gulf currents can affect the climate of coastal regions when the winds pass across. Warm/cold currents heat/cool off the air temperatures over the gulf and bring higher/lower temperatures over land.

With respect to precipitation, the results show that the decrease in annual precipitation is more remarkable in the northern stations (Sulaymaniyah, Salah ad Din, and Zakho) than those in the southern part of the country. The greatest decrease in annual precipitation was observed in Sulaymaniyah at the end of 2080s under RCP8.5. While, the lowest decrease was detected in Najaf station. Although the general trend tends towards a decrease in precipitation across the country, some stations like Kirkuk and Al-Hay show a slight increase in the annual precipitation during the 2020s and 2040s, respectively. Such unforeseen trend is physically uninterpretable especially that most of the surrounded stations show decreasing of the same periods. However, the uncertainty in measurements and modeling results could be the essential reasons behind.

In arid areas, positive changes in temperature will undoubtedly accelerate the process of desertification (Abbasnia et al. 2016) which in turn will ultimately affect on the agricultural activities in the Mesopotamia area. Therefore, this fact highlights the importance of swift adaptation and mitigation measures in the study area. The findings from this study are consistent with many other regional climate changes studies, e.g., in Iran (Najafi and Kermani 2016; Abbasnia et al. 2016), in UAE (Elhakeem et al. 2015), and in Kuwait (Al-Dousari 2005; Al-Awadhi 2014) where they reported an increase in temperature and dust storms and a decrease in precipitation. This was noted in the micro-features printed on surface of drifted sand particles from Iraq towards Kuwait (Al-Dousari and Al-Hazza 2013).

Conclusions

This study investigated the future projection climate change impacts on precipitation and temperature in several stations of Iraq. The NCEP/NCAR predictors were employed to calibrate and validate the SDSM along with the observed local station data. The CMIP5-CanESM2 outputs for the RCP8.5, RCP4.5, and RCP2.6 emission scenarios were used to projects the future changes. Based on the statistical metrics used in this study, results proved the superior performance of the SDSM in modeling present meteorological data. The downscaled results of the GCMs based on the calibrated SDSM showed that temperature would continue to increase in the future time slices. However, future precipitation will increase with more variability and uncertainty across the selected stations, but with consistent trend among RCPs. Overall, annual precipitation in CanESM2 will apparently decrease in the future. There is a clear trend of an increase in temperature and a decrease in precipitation in the study regions. Hence, the findings obtained from this study can be of use to help policy makers in making decisions and planning for adaptation of impacts of climate changes. Moreover, the results can provide a support for better water resources management in Iraq. It is planned to use the outcomes of this study to investigate the impacts of climate changes on water resources availability in Iraq in a further study.