It is in the western Turkey's basins where the scarcity of water, irregular hydrology and great water demands cause droughts to have significant economic, social and environmental consequences. Drought is a complex phenomenon that has a negative agricultural yield impact on the Büyük Menderes river basin, the largest basin in western Turkey, owing to the vulnerability of the region's agricultural sector to weather variability. Increased temperatures leads to an intensification of the hydrological cycle, resulting in dryer dry seasons and wetter rainy seasons, and subsequently heightened risks of more extreme and frequent floods and drought in the basin (Tunalıoğlu and Durdu, 2011). Variability in climate parameters also has significant impact on the availability of water, as well as the quality and quantity of water that is available and accessible in the region. In the Büyük Menderes basin, the most severe drought was recorded in 1947. A recent wave of drought episodes was experienced throughout western Turkey during 2001 and 2002. The drought of 2001 was relatively short in duration (covering the months of April, May and June) but has an important effect, primarily on the corn yield. The last drought event of 2006–2007 clearly exhibited that prolonged periods of rainfall deficit, combined with extremely high summer temperatures, affected the crop production in the basin (Durdu, 2010a, 2010b).
Droughts are caused by scarcity of water and have adversely affected agricultural, water supply and hydropower sectors on a scale ranging from regional to global. In any such sector in a region, a drought can be defined on the basis of a functional relationship between two key factors, the estimated water demand and the expected water supply. The determination of these key factors is affected by the agro-climatic and economic conditions of the region (Kumar and Panu, 1997). Owing to large regional variations in these conditions, drought can be classified into four categories: meteorological, hydrological, agricultural and socio-economic droughts. In this study, the agricultural drought was mainly emphasized for the analysis. A research methodology developed by Mavromatis (2007) and Quiring and Papakryiakou (2003) was adopted in this study. Agricultural drought was defined as an interval of time, generally of the order of months or years, when the moisture supply of a region consistently falls below the climatically appropriate moisture supply such that crop production or range productivity is adversely affected (Quiring and Papakryiakou, 2003). To describe the many scales of drought, a large number of indices were developed in the literature. Drought indices can be useful tools for providing information for decision-makers in business and government, and also to public stakeholders. The most commonly used indices include the Palmer Drought Severity Index (PDSI), its moisture anomaly index (Orig-Z) (Palmer, 1965), the self-calibrated Palmer Drought Severity Index (SC-PDSI), its moisture anomaly index (SC-Z) (Wells et al., 2004), the Standardized Precipitation Index (SPI) (McKee et al., 1993), the NOAA Drought Index (NDI) (Strommen et al., 1980; Titlow, 1987) and the Normalized Difference Vegetation Index based on Vegetation Condition Index (Kogan, 1995). Given the range in derivations and the different responses of these drought indices, not all are suitable for measuring agricultural drought (Quiring and Papakryiakou, 2003).
Numerous studies have been employed the drought indices to evaluate climate change effects on crop production. Tubiello et al. (2002) projected US agricultural production in 2030 and 2090 at 45 representative sites, using scenarios of climate change, developed with the Hadley Centre Model and the Canadian Centre Climate Model and the DSSAT (Decision Support Systems for Agro-technology Transfer) dynamic crop-growth models. Quiring and Papakryiakou (2003) was conducted a comparative performance analysis on four agricultural drought indices to determine the most appropriate index for monitoring agricultural drought and predicting Canada Western Red Spring wheat yield on the Canadian prairies. Chipanshi et al. (2003) was studied the vulnerability of the maize and sorghum crops to climate change were studied using crop simulation models while climate change scenarios were generated from Global Circulation Models. Richter and Semenov (2005) evaluated the impact of climate change on drought indicators and yield of winter wheat in England and Wales using crop simulation model Sirius. Wu et al. (2004) developed an agricultural drought risk-assessment model for Nebraska, USA for corn and soybeans on the basis of variables derived from the SPI and crop-specific drought index using multivariable techniques. Trnka et al. (2007) investigated the relationship between trended district yields of spring barley (1961–2000) and meteorological drought by using Palmer Z-index. In the western Turkey, regional drought indices for evaluating future crop production have not been studied yet. There are several specific drought indices studies to evaluate drought intensity. Özgürel et al. (2002) used the PDSI to determine the periods of drought in western Turkey. Pamuk et al. (2004) employed the SPI to analyse the drought formations in western Turkey.
The future climate projections are important to investigate drought effects on climate-sensitive sectors and to evaluate the potential impacts of climate change at various scales on crop production to reduce their vulnerability. IPCC (2007) indicates that the effects of climate change and increased atmospheric CO2 are expected to lead to decreases in the Mediterranean, the south-west Balkans and in the south of European Russia crop productivity. In southern Europe, general decreases in yield (e.g., legumes − 30 to + 5%; sunflower − 12 to + 3%and tuber crops − 14 to + 7% by 2050) and increases in water demand (e.g., for maize + 2 to + 4% and potato + 6 to + 10% by 2050) are expected for spring sown crops. Therefore, it is important to combine the crop yields with scenario integrations of Regional Climate Models (RCM) for assessing the magnitude of future climate change on crop production.
The main objectives of this study are (1) to evaluate the SPI, the PDSI and its moisture anomaly index Orig-Z and the SC-PDSI and its moisture anomaly index SC-Z based on corn (Zea mays L.) yield departures in the four crop regions (Aydın, Denizli, Afyon, Uşak) in the western Turkey and (2) to investigate the vulnerability of corn production to climate change for Aydın and Afyon regions, using the most appropriate drought index while climate projections have been provided by the Hadley Centre for Climate Prediction and Research ENSEMBLES project (HadCM3Q0).
2. Materials and methods
2.1. Study area
The study area is the Büyük Menderes basin located in the western part of Turkey (Figure 1). The basin is a plain region surrounded by the Aydın Mountain Range, Mount Çökelez and Mount Beşparmak in the north, the West Menteşe Mountain Range, the East Menteşe Mountain range, Mount Karıncalı, Mount Akdağ, Mount Eşler and Mount Söğüt in the south, and the coast of the Aegean Sea in the west. The elevation ranges from the sea level at the western coastal area to more than 2400 m at the southern and northern mountains areas, and the mean elevation of the region is nearly 850 m. At the eastern side of the Büyük Menderes basin the climate is continental; the winters are cold and the summers are hot and the temperature difference between two seasons is large. At the western side of the basin, the climate is typical Mediterranean climate; the summers are usually hot and dry, and in July and August temperatures can reach 43 °C. Mean annual precipitation over the whole basin is about 635 mm and it is distributed unevenly in space and time (Durdu, 2010a, Durdu, 2010b). The mean annual precipitation varies from about 350 mm at the eastern plains to more than 1000 mm at the western mountain areas. Generally rainfall is rare from June to August. Büyük Menderes basin is one of the most productive agricultural regions of Turkey. The main crops cultivated in the basin's plain area are maize, cotton and wheat whereas, olive trees, figs, oranges, peach and plums are cultivated at the foothills of the mountains areas. The Büyük Menderes river and its tributaries traverse the plain area, and the basin total drainage area is about 25 000 km2. The waters of Büyük Menderes River are used primarily for irrigation and domestic use. The intense and extensive cultivation of water demanding crops has lead to a significant water demand increase, which is usually fulfilled by the overexploitation of groundwater resources. The overexploitation of the groundwater, especially during extended dry periods, has lead to the corruption of the already disturbed water balance and the degradation of water resources.
2.2. Palmer Drought Severity Index (PDSI) and its moisture anomaly index (Orig-Z)
The PDSI were developed by Palmer (1965) as a means of measuring the severity of drought. The PDSI method begins with a water balance (usually on a monthly basis) using historic record of precipitation and temperature. Soil moisture storage is handled by dividing the soil into two layers and assuming that 25 mm of water can be stored in the surface layer. The underlying layer has an available capacity that depends on the soil characteristics of the site being considered (Alley, 1984). Moisture cannot be removed from the underlying later until all of the available moisture has been removed from the surface layer. Potential evapotranspiration is calculated using the Thornthwaite (1948) and evapotranspiration losses from the soil occur if potential evapotranspiration is bigger than precipitation. Evapotranspiration loss from the surface layer is assumed to take place at the potential rate and the loss from the underlying layer depends on initial moisture content in the underlying layer, potential evapotranspiration and the combined available moisture capacity in both soil layers. More detailed information about the PDSI calculation procedure can be found in Alley (1984) and Heim (2000).
The Orig-Z index reflects the departure of the weather of a particular month from the average moisture climate for that month regardless of what has occurred in prior or subsequent months (Heim, 2002). The first step in calculating the monthly moisture status (Orig-Z index) is to determine the expected evapotranspiration, runoff, soil moisture loss and recharge rates based on at least a 30-year time series. A water balance equation is subsequently employed to derive the expected or normal precipitation. The monthly departure from normal moisture is determined by comparing the expected precipitation to the actual precipitation. Detailed calculation procedure can be obtained from Quiring and Papakryiakou (2003) and Alley (1984).
2.3. Self-calibrated PDSI (SC-PDSI) and its moisture anomaly index (SC-Z)
The SC-PDSI and its moisture index SC-Z were presented and evaluated by Wells et al. (2004). The model automatically calibrates the behaviour of the index at any location by replacing empirical constants in the index computation with dynamically calculated values. The SC-PDSI calculation procedure is different from the original Palmer index but it does not stray from Palmer's objective. The SC-PDSI replaces the empirically derived climatic characteristics and duration factors with values automatically calculated based upon the historical climatic data of location. The index model establishes separate duration factors for wet and dry spells. The index duration factors are computed using the least squares method for both extremely wet and extremely dry conditions. Theoretically, the SC-PDSI could be calibrated to any category of drought and/or wet spell. The detailed computation steps could be found in Wells et al. (2004).
2.4. The Standardized Precipitation Index (SPI)
The SPI was developed by McKee et al. (1993, 1995) to detect drought and wet periods at different time scales, an important characteristic that is not accomplished with typical drought indices. It is computed by standardizing the probability of observed precipitation for some duration, durations of weeks or months can be used to apply this index to agricultural interests, and longer durations of years can be used to apply this index of water supply and water management interests (Guttman, 1999). The SPI calculation process begins with building a frequency distribution from precipitation data at a location for a specified time period. A gamma probability density function is fitted to the precipitation data and the cumulative distribution of precipitation is determined and an equiprobability transformation is then made from the cumulative distribution to the standard normal distribution with a mean of zero and variance of one (Wu et al., 2001). This transformed probability is the SPI value, which varies between + 2.0 and − 2.0, with extremes outside this range occurring 5% of the time (Edwards and McKee, 1997). Calculation of the SPI requires that there is no missing data in the time series. The data record length is required to be at least 30 years. Guttman (1999) tested different probability models from which the SPI values are computed and found that the Pearson Type III distribution best fits precipitation data. One of the strengths of the SPI is that users can choose the time scale most appropriate for their particular application to compute the SPI. The SPI can be computed for any time period from 1, 2, 3,…, 48,…, to 72 months (Edwards and McKee, 1997).
2.5. ENSEMBLES Global Circulation Model (HadCM3Q0)
In this study, climate scenario data are projected from the HadCM3Q0 GMC, future greenhouse gas emission scenarios and the regional climate model (CLM). The HadCM3Q0 (Collins et al.2006) model is forced by an effective greenhouse forcing corresponding to the provisional IPCC SRES A1B scenario (Nakicenovic and Swart, 2000), are used for the assessment of climate change impacts on future corn production. The HadCM3Q0 model uses flux that the sea surface temperatures (SSTs) remain close to climatological values during a control period, while allowing SSTs to vary from natural variability and from atmospheric forcing, such as increasing CO2 and includes an atmospheric sulfur cycle (Zhang et al., 2012). In this study, a single 1951–2100 integration of the HadCM3Q0 was used for the analysis. The period of 1963–2007 was chosen as current climate and the period of 2056–2100 was used as future climatic projections. The regional climate model CLM is the climate version of the operational limited-area model ‘Lokal-Modell’ (Steppeler et al., 2003), recently renamed to ‘COSMO-model’, of the German Meteorological Service (DWD). General characteristics of the CLM are described in Böhm et al. (2006). The CLM is run at a horizontal grid spacing of 1/6°, corresponding to a grid size of approximately 18 km. In the vertical, 20 atmospheric and 9 soil layers are used (Bachner et al., 2008).
This study was conducted four crop regions, Aydın, Denizli, Afyon, Uşak, located in the Büyük Menderes basin. These regions were chosen due to the availability of high-quality weather data from the meteorological stations and extensive agricultural activities. Monthly air temperature and precipitation data needed for the calculation of drought indices for the period of 1963–2007 were obtained from the State Meteorological Organization and NCDC (2009). Table I demonstrates a descriptive statistics of the temperature and precipitation data. The available water capacity (AWC) values, a location-dependent input parameter required by the models, are obtained from Dağdelen and Gürbüz (2008) and local agricultural organizations.
Table I. Statistical properties of temperature and precipitation data during the period of 1974–2009
Temperature ( °C)
The Büyük Menderes basin experienced severe, extreme and persistent droughts during the period from mid to late 1970s, the period from late 1980s to early 1990s and the hydrological year 2006–2007. The average cumulative areal precipitation during these severe drought periods is compared to normal cumulative areal precipitation in Figure 2. During these three periods, the monthly and annual precipitations were remarkable below normal. Especially, the hydrological years 2006–2007 and 1989–1992 are the first and second driest hydrological years in record, respectively. The prolonged and remarkable decrease of monthly and annual precipitation has a significant impact on water resources of the basin. Usually, the dry periods are associated with high temperatures, which lead to higher evapotranspiration rates and dry soils. These parameters inversely affect both the vegetation and the agriculture of the region as well as the available storage of the reservoirs. Severe and extremely dry conditions lead to irrigation cutbacks, overexploitation of groundwater and dramatic losses of crop yields.
The corn (Zea mays L.) yield data for Aydın, Denizli, Afyon, Uşak regions for the period of 1963–2007 are obtained from the Turkish Statistical Institute and local agricultural organizations. In this study, corn is chosen as a study crop because it has been grown extensively in four regions. In order to eliminate the upward linear trend found in yields owing to advances in agricultural technology as the years progressed, the yield data is detrended by regressing the annual yield against the year of harvest for each region (Mavromatis, 2007) (Figure 3). Furthermore, the residuals of the detrended yield were obtained because the residual variation reflects the effects of weather on yield, and the residuals amplify yield departures from normal, making the variability of yield more obvious (Wu et al., 2004). A positive yield residual indicate the yield is over the multiple year average yield, whereas a negative residual indicate the yield is below the average yield. The unstandardized residuals (hereafter referred to as yield departures) from the detrending procedure are calculated for each crop region that were used in the evaluation of yield models.
In this study, at first, growing season drought indices-based yield models were developed from the statistical data (meteorological data and detrended yield), and then the developed model output was used to compare with the yield residuals. The evaluation of the drought indices for predicting corn yield was carried out in two stages (Quiring and Papakryiakou, 2003; Mavromatis, 2007). During the first step, a growing season drought index was created for each of the five drought indices by summing up the monthly values of indices for June, July and August. Although several other drought index variations were tested, including those related to antecedent moisture conditions during the April (prior to planting), none of these variations significantly improved corn yield. These results are similar to those of Sezgin et al. (2001) who indicate that corn yield is largely influenced by moisture stress during the tassel formation and silking periods which usually occur in June, July and August in the western Turkey. The five growing season drought indices were used to develop a set of curvelinear regression-based corn yield models for each crop region/drought index combination (4 crop regions × 5 drought indices = 20 unique corn yield models). Each regression model was formulated using a second-order polynomial fit to the data, where the independent variable was a growing season drought index (either PDSI, Org-Z, SC-PDSI, SC-Z or SPI) and the dependent variable was the yield departure for that crop region. Equation (1) demonstrates drought indices-based corn yield model curvelinear equations for Aydın crop region:
where Y is yield departures and X is growing season drought index. A second-order polynomial regression function was employed to obtain the corn yield models for the reason that it mindfully estimates the nature of the crop yield water relationship (Ash et al., 1992). In the second step, to figure out the best agricultural drought index, the developed corn yield models were evaluated using four goodness-of-fit measures: the coefficient of determination (R2), the root mean square error (RMSE), the mean absolute error (MAE) and the index of agreement (d). This concept serves an objective means of evaluating the drought indices, which is steady with the definition of agricultural drought and that has superiority in the literature (Kumar and Panu, 1997; Quiring and Papakryiakou, 2003; Mavromatis, 2007). The coefficient of determination (R2) explains the proportion of the total variance in the observed data that can be explained by the model. An R2 value of 0.8 describes that the model explains 80% of the variability in the observed data. R2 value is unsuitable for quantitative analysis because of the outliers in the data. Therefore, it is used with two measures of error (RMSE and MAE) and one measure of model fit (d) to evaluate model performance. The RMSE, combined measure of bias and variance, is given as
The MAE is defined as
The index of agreement (d) is given as
where Pi and Oi are the predicted and observed data, respectively, Ō is average observed data and n is the number of data. The index of agreement measures the degree to which the observed data are approached by the predicted data and ranges from 0 to 1.
In this study, the performance of the crop-yield models was evaluated based on high-drought risk and low drought risk years. Therefore it was prerequisite to determine a criterion for predicting severity of droughts in the crop regions. Based on the most sensitive period of corn growth stages to moisture deficit, it was presumed that the growing seasons with both yield residuals and rainfall amount of June, July and August in the lowest 30% of their probability distribution designate high-risk years.
To investigate the validity of HadCM3Q0 model output, the monthly simulated temperature and precipitation data for 1963 to 2007 (45 years) were compared to the correspondent observed data series with respect to the mean and year-to-year variability. It was assumed that the grid-cell center closest to the each region's meteorological stations represented the conditions at the region (Mavromatis, 2007). To obtain the yield under future scenarios, the most appropriate drought index-based yield model was fed with future climate predictions, and then the model output was compared with the yield residuals. Subject to variations between HadCM3Q0 model output and observations, the crop yield responses (using the most appropriate drought index) to climate change were analysed by producing climate change projection obtained from modelled output of 2056 to 2100 (45 years).
3. Results and discussion
3.1. Determination of the best drought index for the crop regions
Based on the high- and low-risk drought years criterion specified above, 11 of the growing seasons in Aydın crop region were qualified as severely dry years. During these years, the corn yield was − 1.084 ton ha−1 lower, on average, than the respective yield for the 45-year period of 1963–2007. Nine of the growing season in Denizli region was entitled as high-risk drought years and the corn yield was 0.788 ton ha−1 lower than the average corresponding yield in these years. In Afyon and Uşak crop regions, ten of the growing season were monitored as severely dry years and the corresponding corn yield was reduced by − 0.978 and − 0.337 ton ha−1, respectively, in comparison to the historic corn yield mean. Table II demonstrates the performance of the crop yield models based on the drought indices. In Aydın crop region, the SC-PDSI index was the best performed model of the five drought indices. This index was correlated with the corn yield departures with a R2 value of 0.751. The index of agreement with a value of 0.788 indicated that the predicted data were approached the observed data with the least amount of error. The PDSI index was slightly behind the SC-PDSI in Aydın region. In Denizli, Afyon and Uşak crop regions, the PDSI was the best efficient drought index for the high-risk drought years. This index explained 70, 71, and 66% of the variability in the observed data of Denizli, Afyon and Uşak regions, respectively.
Table II. Comparative analysis of the crop yield models based on drought indices
Median (ton ha−1)
SD (ton ha−1)
RMSE (ton ha−1)
Residuals of corn yield for Aydın, n = 11, mean of residuals = − 1.084 ton ha−1
Residuals of corn yield for Denizli, n = 9, mean of residuals = − 0.788 ton ha−1
Residuals of corn yield for Afyon, n = 10, mean of residuals = − 0.978 ton ha−1
Residuals of corn yield for Uşak, n = 10, mean of residuals = − 0.337 ton ha−1
Also, the predicted data agreed very closely with the observed data because the models produced an index of agreement value of 0.782, 0.836 and 0.964 for Denizli, Afyon and Uşak regions, respectively. The SC-PDSI indices in these regions were not performed well with slightly poorer summary statistics. Overall in the crop regions, the moisture anomaly indices Orig-Z and SC-Z were equally efficient in most model analysis. The reason for that the calculated climate characteristic parameters, based solely on the climate of the location, for each region were close to each other. The detailed computation of climate characteristic parameters can be found in Wells et al. (2004). Summary statistics of the moisture anomaly indices were close to the corresponding data of the SC-PDSI. The SPI index was not performed well for evaluation of corn production in the crop regions. In the fact that the SPI depends on monthly precipitation values, the PDSI indices consider evapotranspiration rates in the calculation processes.
The crop yield models that ranked best at high-risk drought years were also performed well at predicting the observed corn yields during the low-risk drought years. In Aydın regions, the predicted data from crop yield model based on SC-PDSI index explained only 28% of the observed data. This value is much poorer than the correspondent value of the high-risk drought years. The model based on the PDSI index explained only 23, 21 and 18% of the total variance in the observed departures of corn yield in Denizli, Afyon and Uşak regions, respectively. Mavromatis (2007) reached the same conclusion when he was evaluating the drought indices for assessing future wheat production in Greece. He indicated that the goodness-of-fit measures were much poorer in the low-risk drought years than the corresponding values for the high-risk drought years. Quiring and Papakryiakou (2003) indicates that when moisture conditions are not limiting like in low-risk drought years, other factors such as soil fertility, the presence or absence of disease/pests, and the amount of fertilized applied, become important determinants of crop yield. In addition, Mavromatis (2007) emphasizes that the drought indices considered in this study do not properly model soil water especially excess rainfall and give too little regard to crop management and phonological drought sensitivities.
Figure 4 demonstrates the best drought index-based models for all years (high-risk and low-risk drought years) by plotting the observed versus modelled yield departures. In Aydın crop region, the crop yield model based on the SC-PDSI index explained 43% of the observed variability in corn yield. Figure 4(a) indicates the crop yield model is a good predictor during the years with negative departures, however, its performance is not well to produce positive yield departures greater than + 1.1 ton ha−1. The crop yield model based on the PDSI index accounted for 41, 43 and 42% of the total variance in the observed yield departures for Denizli, Afyon and Uşak regions, respectively (Figure 4(b)–(d)). Similar to the SC-PDSI-based model, the PDSI-based yield model produced underestimated yield departures during the years with large corn production. The PDSI-based yield models did not perform well to estimate positive yield departures greater than + 1.2, + 0.7 and + 0.38 ton ha−1 for Denizli, Afyon and Uşak regions, respectively.
Assessment of these results indicates that the Palmer indices are the most appropriate for predicting corn yield departures during the high-risk drought years in which the moisture stress is a significant factor on yield amount. In low-risk drought years in which the water is not scarce resource, these indices are not performed well for predicting yield departures. Mavromatis (2007) found the same results when he evaluated seven drought indices for predicting future wheat yield in two crop regions in Greece which is 300 km away from the Büyük Menderes basin. He indicated that the best performed models based on the drought indices were the PDSI and SC-PDSI. These indices were having a tendency to underpredict positive yield departures. In his study, the moisture anomaly indices did not perform well as much as the PDSI indices did. However, when Quiring and Papakryiakou (2003) were studying the performance of four agricultural drought indices [PDSI, Orig-Z, SPI and NOAA Drought Index (NDI)] for prediction of spring wheat on the Canadian prairies, in contrast to Mavromatis (2007), they found that the Orig-Z clearly outperformed the PDSI index and was the best performed index for measuring agricultural drought on the Canadian prairies. Quiring and Papakryiakou (2003) also indicated that the SPI index was ranked as second appropriate index after the Orig-Z, whereas, Mavromatis (2007) found that the SPI was the least successful index for assessing wheat production. In this study, the results are similar to Mavromatis (2007) findings in which the Palmer drought indices were proved to be the best indices for predicting wheat yield in comparison to the SPI index. The reason for that the Palmer's approach is more physically based than the SPI index and it considers potential and actual evapotranspiration, which are important variables in determining crop growth (Quiring and Papakryiakou, 2003; Mavromatis, 2007). In addition, the SPI calculation procedure depends on precipitation, whereas corn yield may also be reduced due to high air temperatures, especially during pollination period (July). When evapotranspiration exceeds water supply from the soil at any time during corn growth, corn yields are reduced. NC Cooperative Extension (2000) indicates that the amount of yield loss that occurs during dry weather depends on what growth stage the corn is in and how severe the dry conditions become. Four days of visible wilting only a week prior to tasseling and the milk stage may reduce yield by 50%or more (NC Cooperative Extension, 2000). Not only precipitation but temperature and environmental conditions are also significant parameters in corn yield production. Therefore, the PDSI-based drought indices exhibited better performance than the SPI did.
Spatial variability is another issue of the performance of drought indices-based yield models. While the SC-PDSI-based yield model was the best successful drought index in Aydın region located coastal part of the Büyük Menderes river basin, the PDSI-based model was good at in Denizli, Afyon and Uşak regions described as highlands of the basin. Quiring and Papakryiakou (2003) indicates that the differences in model performance from one crop district to another may occur due to regional variations in climate, soil conditions, response to moisture stress or data quality. In addition, although corn yield is vulnerable to growing season moisture conditions, there are many other components, such as disease, storm damage, soil conditions and pests, that reduce corn yield. As the yield models based on the SC-PDSI in Aydın region and the PDSI in Denizli, Afyon, Uşak regions were performed best in predicting the observed corn yield data, they will be employed for all subsequent work involving drought indices.
3.2. Assessment of crop yield responses with future climate projections
Before running into the evaluation of crop yield responses to climate change, the data for the period of 1963–2007 was compared to the same period of the observed data for validation of the model output. Owing to the limitation of the model resolution, the crop yield departures were evaluated for Aydın and Afyon regions. Figure 5 demonstrates the differences in temperature on monthly basis between the HadCM3Q0 output and observed data for Aydın and Afyon crop regions. The model output indicates that Aydın region is colder for September–April period but is warmer for May–August period in comparison to the observed data (Figure 5(a)). The model outputs for Afyon region demonstrate warmer conditions for the period of May–November (Figure 5(b)). The model outputs for annual average temperature overestimate the observed data by 0.6 °C in Aydın (warmer) region and underestimate 1.0 °C in Afyon (colder) region. These results were similar to Mavromatis (2007) in which he found that the simulated temperatures on annual basis underestimated the observed values by 1.2 °C in northern region (colder) and 1.0 °C in southern region (warmer).
The modelled monthly basis precipitation data and observed data for the period of 1963–2007 are compared in Figure 6. The model outputs were evaluated based on the estimates of the coefficient of variation (CV), expressed as SD*100/mean. Mavromatis (2007) suggested that the estimates of CV of the modelled data should be within about 30% of the corresponding values of the observed data. In Aydın region, a good agreement, particularly during the critical months June–August, was obtained between the observations and the HadCM3Q0 simulated monthly precipitations (Figure 6(a)). Statistically significant differences in precipitation occurred in January, February and March for Aydın region. However, the agreement between simulated and observed precipitation data was not well in Afyon region (Figure 6(b)). There were three cases (March, April and May) of statistically significant differences in precipitation in Afyon. Jones et al. (2003) in his paper indicates that agreement between observed data and simulated model output for a particular period might be improved/worsened if the climate model results for other model years encompassing the different period are used. The simulation by HadCM3Q0 is just one integration of the climate model. Different initial conditions would have resulted in a different simulation for the 1963–2007. Differences would be relatively small at the global scale but can be much larger at the grid-box scale of these comparisons (Jones et al., 2003). Each model grid-box is at an elevation equal to average for that region in the real world, a very simplified form of the true orography. Slight difference from true orography will particularly influence temperature, precipitation and radiation (Jones et al., 2003).
Although, the simulated model output and observed data for temperature and precipitation have an acceptable closeness, direct application of output from the model to produce drought indices is often, despite some exceptions (Mavromatis and Jones, 1999), inadequate because of the limited representation of mesoscale atmospheric processes, topography and land-sea distribution in climate models. Moreover, climate models exhibit a much larger spatial scale (grid-point area) than is usually needed in impact studies and this leads to inconsistencies in frequency statistics (Schmidli et al., 2007). Mavromatis (2007) indicated that the differences between the monthly PDSI series with observed and simulated inputs were sufficiently large that use of HadRM3 climate model output for 2071–2100 directly with the drought indices would lead to the dubious future responses about crop yield. Therefore, in this study a common method of dealing with climate model inadequacies known as the delta change method was used to compute difference between current and future climate simulations and add these changes to observed data. Applying the delta change method assumes that climate models more reliably simulate relative changes rather than absolute values (Hay et al., 2000). The downscaling formulation for temperature and precipitation were formulated as follows.
where i and j stand for ith d of the jth month, Tm is temperature data produced from raw climate mode output, Tobs is observed data, ΔT(j) is the change in temperature for month j (Figure 5(a) and (b)), T̄m(2056−2100) is modelled average raw temperature data for future projections, T̄m(1963−2007) is modelled average raw temperature for current time period, Pm is precipitation data produced from raw climate model output, Pobs is observed precipitation data, ΔP(j) is the change in precipitation for month j (Figure 6(a) and (b)), P̄m(2056−2100) is modelled average raw precipitation data for future projections, and P̄m(1963−2007) is modelled average raw precipitation for current time period.
Once the raw simulated model output was downscaled as future scenario data, the SC-PDSI and PDSI drought indices were produced for Aydın (warmer) and Afyon (colder) regions, respectively. The crop yield model based on the SC-PDSI index was combined with future scenario data to evaluate the variations in future crop yield. The corn yield residuals in Aydın region seriously dropped by 2.1 ton ha−1, on average, indicating extremely reduced output in high-risk drought years (Table III). Future scenario data with crop yield model produced slightly lower losses, around − 0.014 ton ha−1, for the high-risk year in Afyon region. In low-risk drought years, in which the moisture stress not being a significant yield-limiting factor, the crop yield models produced a positive yield response of 0.022 ton ha−1 in Aydın region. However, the yield models predicted lower yield by 0.104 ton ha−1 for corn in Afyon region (Table III).
Table III. Comparative assessment of modelled yields and observed data for future high and low drought risk years
Mean (ton ha−1)
SD (ton ha−1)
The crop trends on future scenario was further attempted by plotting the modelled yield departures for all years using the baseline climate versus the corresponding climate projections based on HadCM3Q0 data (Figure 7). In Aydın region, the future climate projections oriented yield model based on SC-PDSI index explained 41% of the observed variability in corn yield. The model was not successful to predict positive yield departures greater than 1.37 ton ha−1 (Figure 7(a)). In Afyon region, the crop yield model based on the PDSI index accounted for 40% of the total variance in the observed yield departures. However, the model was not good at predicting the positive yield departures greater than 0.72 ton ha−1 (Figure 7(b)). It is obvious in Table III for Afyon region (altitude is 1100 m), in high-risk drought years, there is a less yield loss in comparison to Aydın (altitude is 50 m). There is evidence from several studies (Penuelas and Boada, 2003; Camarero and Gutierrez, 2004) that indicates that temperatures are increasing rapidly in high-elevation mountain regions (IPCC, 2007). This might create appropriate conditions for corn growth in colder Afyon region.
Overall, this study indicates that the adverse effects of a warmer and drier climate change may exhibit different impacts depending on location and crop type. In high-drought risk years, the drop in yields for future climate scenarios in Afyon region is relatively smaller compared with the Aydın region (Table III). The relatively higher yield reductions in Aydın region point to endless of problems that currently annoy this region in light of the high evaporative losses and high permeability of soil. As a result growing a water demanding crop in Aydın regions such as corn under a drier climate may become even more problematic. The published IPCC reports are strengthening these findings of this study. IPCC (2007) indicate that increasing temperatures and shifting rainfall sequence will change from one agro-ecosystem to another. The report also emphasizes that mean annual precipitation is projected to decrease along the Mediterranean and in south-eastern Europe. Differences in water availability between regions are anticipated to become sharper and annual average runoff will decrease in the south-eastern Europe. Furthermore, crop suitability is likely to change throughout Europe, and crop productivity is likely to decrease in south-eastern Europe (IPCC, 2007). In particular, in the Mediterranean region, increases in the frequency of extreme climate events during specific crop development stages (rainy day during sowing time, heat stress during flowering period), together with higher rainfall intensity and longer dry spells, are likely to reduce the yield of summer crops (IPCC, 2007; Mavromatis, 2011). Dixon et al. (1994) indicated that measuring weather-related factor by growth stage was particularly important for predicting yields under climate change.
The limitations of this study depend on choosing the most appropriate drought index and local agro-climatic conditions. Determining the best suited drought index for a crop region is particularly difficult because the answer will vary depending on the crop's sensitivity to moisture shortage and the characteristics of the study region, including soil properties and climate regime (Mavromatis, 2007). Some drought indices combine the effects of temperature and precipitation in drought monitoring, while others are based solely on the precipitation data. Therefore, different kind of indices should always be tested to determine the best suited drought index for a particular application.
Drought indices-based yield models to predict corn yield can be useful tool for providing information for decision-makers and to public stakeholders. As the response of crop yield models will vary depending on the crop's sensitivity to moisture shortage and the properties of the crop region, such as, soil characteristics, climate regime and economic factors, it is difficult to determine the most appropriate drought index. In this study, while the crop yield model based on the SC-PDSI was the best performed model in Aydn region, the PDSI was best performed model in Denizli, Afyon and Uşak crop regions in high-risk drought years. These crop yield models were also the most appropriate for predicting the observed wheat yields in the years when moisture was not a significant limiting factor. However, the correlation between the drought indices and crop yield models was not strong. Mavromatis (2007) improved potential evapotranspiration term in the SC-PDSI model but this did not develop the model's performance.
The comparative analysis between the observed monthly precipitation and temperature data and those derived from the HadCM3Q0 climate model shows an acceptable agreement. Owing to the resolution of the HadCM3Q0 climate model, only Aydın (warmer) and Afyon (colder) regions were compared to analyse the future climate scenario oriented yield models based on the drought indices. Based on the future climate projections developed, the yield departures in Aydın region dramatically dropped by 2.1 ton ha−1 in high-risk drought years. Clearly, in Aydın highly negative trends in yield are expected with the climate change scenario in the growing seasons, with severe moisture deficit. High-drought risk years have hot daytime temperatures in July and August. Because the high daily temperature causes plant respiration and high respiration results in large grain loss, high temperature in July and August affect corn yield negatively. Greater temperature variability may cause to more variable soil moisture and crop yield departures, and larger soil moisture deficit and crop yield reduction are likely to eventuate more frequently. As corn growth stages are so sensitive to moisture stress, measuring weather-related factors is particularly important for predicting yields under climate change. It should be noted that possible changes in sowing data and hybrid selection can reduce negative impact of potential warming on corn yield in Aydın. There was slight yield reduction in Afyon region with a value of 0.014 ton ha−1 in high-risk drought years. In low-risk drought years, the crop yield models produced a positive yield departure of 0.022 ton ha−1 in Aydın region; however, it generated a negative yield departure of 0.104 ton ha−1 in Afyon region. Further research and data are needed to clearly identify the usefulness of drought indices on corn yield models. Especially high-resolution climate models will develop the robustness of the models.