There are many environmental challenges in water-limited places in the 21st century, particularly in dry and semi-arid regions, due to the threat of climate change caused by the greenhouse effect. This study intends to explore and assess the influence of climate change on hydro-climatological parameters using statistical downscaling and future forecasts of mean monthly precipitation and temperature throughout Famagusta (Mağusa), Nicosia (Lefkoşa), and Kyrenia (Girne) stations, North Cyprus. To achieve the study's goal, 13 predictors of BNU-ESM GCMs from CMIP5 were used at a grid point in the Karfas region. To find the primary predictors, GCM data were screened using mutual information (MI) and correlation coefficient (CC) feature extraction methods prior to downscaling modeling. A neural network (ANN), an adaptive neuro fuzzy inference system (ANFIS), and multiple linear regression (MLR) models were employed as the downscaling models. We used the best downscaling model as a benchmark for future precipitation and temperature estimates for the period 2018–2040 under the RCP4.5 scenario. In the future, Famagusta and Nicosia would have up to 22% less rain, and Famagusta and Kyrenia will have 2.9% greater heat. The findings of this research could be useful in decision-making, as well as water resource management and climate change.

  • The paper use the downscaling modeling for parameter evaluation.

  • It reveals the temperature and rainfall variation for future prediction, 2045.

  • It analyses the effective parameter/s used for present and future evaluation.

  • It is a good reference source for researchers dealing with the topic.

Water resources and many other life aspects are affected by the global change of climate in many developing and developed countries around the world (Alotaibi et al. 2018). Over the recent decades, research on climate change has achieved tremendous significance due to global warming. In several parts of the globe, the variations in rainfall and temperature differs owing to the consequence of climate changes (Peng et al. 2018).

The general circulation models (GCMs) are used to study the future fluctuations of temperature and precipitation in such a way, that up to the end of the century, large-scale climate data are simulated under the greenhouse gas changes effect. In the local climate studies, the developed GCM outputs in coarse spatial resolution cannot directly be used. As a result, suitable techniques should be used to downscale the coarse spatial resolution of GCM outputs into finer local climate data (Mora et al. 2014). The downscaling models are generally classified into: (a) dynamical downscaling method that uses regional climate models (RCMs) based on boundary condition set by GCM data to extract local scale information and (b) statistical downscaling method which is based on creating statistical relationship between the predictand (local scale weather data) and predictors (large-scale climate variables). The methods used for statistical downscaling are simple to use and require little efforts for computations and can be applied in different regions for different GCM outputs (Timbal et al. 2003). Many methods of statistical downscaling can be found in the literature, including statistical downscaling model (SDSM) (Samadi et al. 2013a, 2013b), correlation analysis (Landman et al. 2001), multiple linear regression (MLR) (Klein 1983), the nearest neighborhood (Zorita & Von Storch 1999), adaptive neuro fuzzy inference system (ANFIS) (Alotaibi et al. 2018), and artificial neural network (ANN) (Nourani et al. 2018) among others, which have already been used to downscale statistically the GCM outputs.

Focusing on the survey of literature on the consequence results achieved for statistical downscaling modeling of climatic parameters based on the application of AI-based methods, e.g., ANNs, it was found that there are contradicting issues regarding their performance; while some studies show efficiency and superiority of ANNs, others demonstrate inferiority and drawback of ANNs over MLR-based models (e.g., Abdellatif et al. 2013). The quantity and quality of the used data may contribute to the inconsistency of these results. AI-based modeling challenging issues for huge amounts of data arise from noise presence in the data set and redundant information. While using the nonlinear models (such as ANN and ANFIS), over the simulation time, the noise present in the data set can be nonlinearly magnified. According to a study by Bowden et al. (2005), in ANN-based hydro-climatic processes modeling, utilization of several input variables may lead to inefficient results owing to: (i) irrelevant input variables, which lead to difficulty in training process; (ii) lack of convergence and low precision can be caused by the irrelevant input variables; (iii) being more time-consuming and an increase in computational memory; and (iv) it is more difficult to understand the complex models formed using huge inputs in comparison to simple models, more especially when the results of the models are compared. As such, the AI-based downscaling models' efficiency can be largely enhanced by input feature extraction methods as a preprocessing technique.

Some studies, such as Babel et al. (2017), utilized data preprocessing techniques in addressing the complexity in statistical downscaling modeling of GCM data through extraction of dominant predictors. However, feature extraction methods are the techniques employed for the selection of the dominant predictors for downscaling modeling with maximum impact. Cheng et al. (2005) applied feature extraction methods for hydro-climatic variables' modeling using fuzzy model and coupled genetic algorithm. Pahlavan et al. (2017) employed feature extraction method via mutual information (MI) for predictors screening for downscaling modeling using SDSM. Okkan (2015) applied data preprocessing methods of MI and correlation coefficient (CC) for variables' input selection. Despite an increase in variation threats caused by climate change on hydro-climatological variables, previous studies show no efforts have been made to ascertain and evaluate the impact that the climate change may have on precipitation and temperature in North Cyprus through statistical downscaling. Therefore, to fill this gap, this study employed the applications of AI-based (ANN and ANFIS) and MLR statistical downscaling techniques for precipitation and temperature projections over Famagusta (Mağusa), Nicosia (Lefkoşa), and Kyrenia (Girne) stations. Among the several methods of data preprocessing for feature extraction of dominant predictors, MI and CC were found to be exceptional, with remarkable performance and, therefore, were used in this study for the selection of dominant predictors.

Famagusta (Mağusa) station

Famagusta is a coastal city, located at a latitude 35.12°N, longitude 33.94°E, and altitude 25 m. The climate of Famagusta is warm and temperate with around 19.3 °C average temperature and around 407 mm average rainfall annually.

Nicosia (Lefkoşa) station

Nicosia is the capital city of both South and North Cyprus. Nicosia is characterized by hot summers and wet winters and, as such, daytime maximum temperature and night-time minimum temperature difference is greater than in the coastal cities. Comparing the other coastline cities' daytime temperature with Nicosia, there is a difference of up to 4 °C or 7 °C for the hottest months of July and August. In the coldest months of January and February, up to 2 °C or 3 °C lower daytime temperature difference is witnessed in Nicosia than other coastal cities.

Kyrenia (Girne) station

Kyrenia is a coastal city in North Cyprus. The city is situated at a latitude of 35.34°N, longitude of 33.32°E, and altitude of 3 m. The average temperature in the hottest month of July is 29 °C and in the coldest month of January 10 °C. Rainfall in hilly areas varies from 750 mm to as low as 500 mm, falls mostly in winter and on rare occasions in summer. Due to elevation, the highest rainfall is experienced in the Kyrenia Mountains at a rate which ranges between 750 and 1,110 mm. Figure 1 shows a map of Cyprus and locations of the study and the selected grid point.

Figure 1

Study locations and the selected grid point.

Figure 1

Study locations and the selected grid point.

Study data

The mean monthly predictands data of precipitation and temperature used in this study from 1995 to 2017 (23 years) were obtained from North Cyprus Meteorological Organization. A total of 13 GCM predictors for a grid point located at Karfas for BNU-ESM GCM under Representation Concentration Pathways (RCPs) from CMIP5 was downloaded from http://cera-www.dkrz.de for large-scale predictors from 1995 to 2040 for the statistical downscaling and future projection of the predictands. Table 1 and Figure 2 give the descriptive statistics of the predictands and precipitation time series from 1995 to 2017 for all stations.

Table 1

Data description for predictands

StationParametersUnitMin.Max.AverageSt. deviation
Famagusta (Mağusa) Precipitation (PRmm 5.33 0.77 0.93 
Mean temperature (Tmean°C 12.37 31.23 21.47 5.88 
Kyrenia (Girne) Precipitation (PRmm 189.69 27.69 33.13 
Mean temperature (Tmean°C 11.16 30.96 20.61 6.11 
Nicosia (Lefkosa) Precipitation (PRmm 5.83 0.94 1.11 
Mean temperature (Tmean°C 10.76 33.26 24.17 6.85 
StationParametersUnitMin.Max.AverageSt. deviation
Famagusta (Mağusa) Precipitation (PRmm 5.33 0.77 0.93 
Mean temperature (Tmean°C 12.37 31.23 21.47 5.88 
Kyrenia (Girne) Precipitation (PRmm 189.69 27.69 33.13 
Mean temperature (Tmean°C 11.16 30.96 20.61 6.11 
Nicosia (Lefkosa) Precipitation (PRmm 5.83 0.94 1.11 
Mean temperature (Tmean°C 10.76 33.26 24.17 6.85 
Figure 2

Precipitation time series from 1995 to 2017 for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

Figure 2

Precipitation time series from 1995 to 2017 for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

From the displayed precipitation time series in Figure 2 it can be deduced that the precipitation patterns for Famagusta, Nicosia, and Kyrenia are similar, which might be due to their identical semi-arid climate. Despite this development, a difference in the amount or rate of precipitation can be seen from the highest precipitation station (Kyrenia) to the lowest precipitation station (Famagusta). This is in agreement with the characteristics of the data shown in Table 1. In spite of the coastal location of Famagusta station, which by virtue of the larger evaporation body (e.g., sea) would experience higher precipitation than the inland (Nicosia) station, the time series data plotted in Figure 2 show more precipitation was received by Nicosia than Famagusta. This might be because both Nicosia and Kyrenia are surrounded by mountains and the energy reaching the Earth's surface which helps in evaporation and eventual precipitation will reach the mountains faster and, as such, may aggravate precipitation which may lead to receiving more rainfall in mountainous areas. Moreover, from the water cycle concept, water is in constant movement from the Earth to the atmosphere and due to such movement, the water that evaporates from a given area may fall in another area. This implies that a large quantity of water that evaporates from Famagusta may not fall in Famagusta, rather, a small quantity of water that evaporates from another area may fall in Famagusta due to the continuous movement of the water in the atmosphere. Figure 3 gives the temperature time series from 1995 to 2017 for all stations.

Figure 3

Temperature time series from 1995 to 2017 for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

Figure 3

Temperature time series from 1995 to 2017 for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

Table 2 describes the GCM data employed for the study.

Table 2

Data descriptions of large-scale GCM predictors

S/NPredictorsDescriptionHeight
Tasmax Maximum air temperature Surface 
Ts Sea surface temperature Surface 
Hurs Relative humidity Near surface 
Huss Specific humidity Near surface 
Psl Sea level pressure Surface 
Evspsbl Water evaporation Surface 
Sfcwind Wind speed Near surface 
Ps Surface air pressure Surface 
Uas Zonal wind speed Near surface 
10 Vas Meridional wind speed Near surface 
11 Prw Atmospheric water vapor content Surface 
12 Pr Precipitation Surface 
13 Clwvi Cloud condensed water content Surface 
S/NPredictorsDescriptionHeight
Tasmax Maximum air temperature Surface 
Ts Sea surface temperature Surface 
Hurs Relative humidity Near surface 
Huss Specific humidity Near surface 
Psl Sea level pressure Surface 
Evspsbl Water evaporation Surface 
Sfcwind Wind speed Near surface 
Ps Surface air pressure Surface 
Uas Zonal wind speed Near surface 
10 Vas Meridional wind speed Near surface 
11 Prw Atmospheric water vapor content Surface 
12 Pr Precipitation Surface 
13 Clwvi Cloud condensed water content Surface 

As seen in Table 2, the predictors used for the purpose of the study include variables that are related to the factors influencing both precipitation and temperature such as wind speed, relative humidity, temperature, precipitation, etc. at surface and near surface heights.

Data normalization and performance criteria

The data in this study were normalized as follows:
formula
(1)
where , , and are the normalized value, ith observed value, maximum value, and minimum value, respectively.
To determine the models' performance, determination coefficient (DC) and root mean square error (RMSE) were employed (Legates & McCabe 1999), given by:
formula
(2)
formula
(3)
where is the observed ith value, represents mean of the observed values, is the predicted ith value, and N stands for number of observations. The DC has values which range between −infinity and 1 (−∞ < DC ≤ 1) and the model is more efficient with a value close to 1 while RMSE lies between 0 and infinity (0 ≥ RMSE < ∞) with the efficiency of the model increasing with RMSE towards 0 (Nourani et al. 2020).

Model validation

In this thesis, k-fold random division of the data set was done for k = 4 (four-fold) subsamples. In this way, three-quarters of the subsample were utilized for training and the one-quarter remaining subsample utilized for model validation. The process was repeated four consecutive times, with different three-quarter and one-quarter subsamples for training and validation.

Proposed methodology

The proposed methodology is in three steps as follows.

Step 1: feature extraction of dominant predictors

For different GCMs, there are different methods of obtaining outputs due to different resolutions; as a result, the reliability of the different GCMs determines the projection efficiency for the future. Therefore, GCM performance assessment and determination of the most dominant GCMs become important (Nourani et al. 2019a, 2019b). In this way, CC and MI feature extraction methods were used to calculate predictands (precipitation and temperature) and predictors (GCM outputs) relationships in the form of linear and nonlinear approaches, which led to the selection of dominant predictors used as inputs for the downscaling modeling.

Step 2: statistical downscaling modeling of precipitation and temperature

The statistical downscaling of precipitation and temperature using ANN and ANFIS models was performed in this step considering the dominant predictors determined by feature extraction methods in step 1.

In accordance with the technical literature survey that was conducted, it was found that the evaluation of the AI-based models' downscaling efficiency has been carried out mostly using classic models such as SDSM. In statistical downscaling modeling, SDSM is a commonly applied method which determines predictant and predictors statistical relevance through utilization of MLR method. Hence, MLR model was also employed for the statistical downscaling of the predictands to compare and evaluate the performance of AI-based (ANN and ANFIS) downscaling models. The reason why MLR was preferred over SDSM downscaling model is that, in any time scale, MLR can use predictors, whereas predictors can only be used in daily scale using the SDSM model. In view of this study's monthly time scale, working with the model that uses only daily time scale is not possible. In the next step (step 3), the downscaling model that produced the highest efficiency was used as a benchmark for the projection of future precipitation and temperature.

Step 3: future projection of precipitation and temperature

The future projection of monthly precipitation sum and mean temperature for the future from 2018 to 2040 was conducted using the best downscaled model as benchmark output. It is important to mention that due to social, population, energy extension and economic growth rate, future anthropogenic functions are admissibly plausible, which have an impact on greenhouse gas production. According to the future projection above, RCPs represent emission scenario and radiative forcing pathways and were used for precipitation and temperature in the future projection. Detailed information regarding RCPs and emission scenarios can be found in Nourani et al. (2019a, 2019b) study. Figure 4 shows the schematic procedure developed for the proposed modeling. It should be noted that the proposed methodology is the same for both precipitation and temperature and is applied for all stations.

Figure 4

The applied study methodology.

Figure 4

The applied study methodology.

Artificial neural network (ANN)

Among AI techniques, ANN is a convincing approach that has the potential to handle noisy, nonlinearity, and data dynamism, more specifically when there is insufficient understanding of the underlying physical relationships. This emphasizes the ANN as an important data-driven tool for time series modeling (Abdullahi & Elkiran 2017; Nourani et al. 2020).

ANN comprises neurons or nodes interconnected by a number of simple processing elements with processing characteristics of attractive peculiarity of information including generalization capability, nonlinearity, learning, parallelism, and noise tolerance. For handling many problems in engineering, a feed forward neural network (FFNN) model equipped with an algorithm called back propagation (BP), among neural network methods, is mostly used (Hornik et al. 1989). The method of FFNN constitutes layers of neurons of parallel processing elements, whereby weights are used to connect fully every layer to the proceeding layer. The BP algorithm is generally used to accomplish such ANNs' learning (Hornik et al. 1989).

Levenberg–Marquardt (LM), among several algorithms for neural networks' training, was preferred in this study owing to its ability to converge quickly (Sahoo et al. 2005). As a hybrid technique and for optimal convergence of solution, the LM algorithm utilizes Gauss–Newton and steepest descent together. The hybrid approach to solve problems of wide margins, usually, different algorithms' best characteristics are used for trade off. For example, the approach of Gauss–Newton is quicker when the optimal solution is relatively close to the initial guess, consequently, the LM algorithm is utilized. Contrarily, the approach of steepest descent is used by the LM algorithm to find an area of potential solution and traverse the design space and then determine the optimum. This technique is especially important for providing solutions to nonlinear equations (Wilson & Mantooth 2013). Nevertheless, the tangent sigmoid (tansig) was utilized as the transfer function of both layers (hidden and output). To ensure hidden layer neurons of optimum number and also to generate efficient epochs for calibration, a process of trial and error was followed.

Adaptive neuro fuzzy inference system (ANFIS)

‘Neuro fuzzy’ in neural network technical literature refers to the approaches of fuzzy logic modeling that comprise distinctive learning algorithm application to the system of fuzzy inference (FIS). In the development of neuro fuzzy system, ANFIS is a peculiar approach that Jang (1993) initially introduced. It utilizes the learning algorithm of the neural network.

Every fuzzy system is constituted of three major parts, namely: fuzzifier, defuzzifier, and fuzzy database (Nourani et al. 2015). Inference engine and fuzzy rule base are fuzzy database's two essential parts. Rules related to fuzzy propositions are contained in fuzzy rule base as Jang et al. (1997) revealed. Therefore, in fuzzy inference, operational analysis is administered. The fuzzy inference engines that are most commonly applied are Mamdani and Sugeno, both of which could produce a desired goal on application.

For overall real continuous functions, ANFIS being a universal approximator is capable of impacting accuracy to a higher degree. As Jang et al. (1997) stated, functionally, ANFIS is equivalent to FIS. Thus, specifically, the ANFIS system is equal to first order Sugeno fuzzy model. Figure 4 shows ANFIS general structure. As seen in Figure 4, the ANFIS model has x and y inputs with f considered as output. There are two if-then sets of rules constituted in the first order Sugeno model, demonstrated as:
formula
(4)
formula
(5)
where x inputs' membership functions (MFs) are represented by A1 and A2, while y inputs' MFs are represented by B1 and B2. p1, q1, r1 and p2, q2, r2 are output function parameters.

Each ANFIS layer has the following functions:

Layer 1: Input variable membership grades are produced by each node in this layer. The ith node output of the k layer is denoted by . Assuming as a generalized bell function of a MF (gbellmf), Jang & Sun (1995) stated that the output (), can be determined from:
formula
(6)
where , , are adaptable variables called premise parameters.
Layer 2: In this layer, the incoming signals are multiplied by each node:
formula
(7)
Layer 3: In this layer, the th node calculated the normalized firing strength:
formula
(8)
Layer 4: In this layer, node i calculated the contribution given by the th rule to the model output:
formula
(9)
where is the perimeter set, is the output of layer 3.
Layer 5: In this layer, a single node calculated the overall output of the ANFIS (Jang & Sun 1995):
formula
(10)

ANFIS hybrid learning algorithm is formed by a combination of least-squares and gradient descent methods. The consequent parameters are represented by while the optimization parameters known as premise parameters are represented by , , . Least-square technique identified the consequent parameter in the hybrid learning approach towards forward pass until the fourth layer where the output node goes forward. There is propagation of the error signals backward in the backward pass and the gradient descent updated the premise parameters (Nourani & Komasi 2013).

Multiple linear regression (MLR)

MLR is a conventional technique that models mathematically the linear relevancy between two or more predictors (independent variables) and a predictand (dependent variable). In general, the n and y representing predictors and dependent variables might have a relation, given by (Nourani et al. 2019a, 2019b):
formula
(11)
where the ith predictor value is , the constant of regression is , the ith predictor coefficient is and the error term is . Figure 4 shows a visual MLR model.

Correlation coefficient (CC)

The method of CC has been applied as a common metric for understanding the linear behavior between variables (predictand and predictors). Its values fall between +1 and −1 (Baghanam et al. 2019).

Mutual information (MI)

Using data probability distribution, mathematically the information and entropy content have been measured and formulated (Shannon 1948).

Model validation

In this thesis, k-fold cross-validation technique was utilized to validate the models. With k-fold technique, the data set is divided randomly into a certain number of equal subsamples regarded as k-number subsamples. The model is trained with the k-number subsamples minus 1 (k − 1) and the remaining k-subsample is used for validation. The process is repeated continuously with k − 1 and k different subsamples for training and validation of the model, until the subsamples are used once in both training and validation of the model. The final results are then obtained by averaging the k-fold results. The advantage of k-fold validation technique is that for model training and validation, every observation is used once (Nourani et al. 2019a, 2019b).

In this study, the k-fold random division of the data set was done for k = 4 (four-fold) subsamples. In this way, three-quarters of the subsample were utilized for training and the one-quarter remaining subsample utilized for model validation. The process was repeated four consecutive times, with different three-quarter and one-quarter subsamples for training and validation.

This study was performed in three steps. In the first step, the most appropriate downscaling predictors were determined using the methods of MI and CC in order to have better downscaling modeling of precipitation and temperature. Statistical downscaling modeling was performed in the second step through the application of ANN and ANFIS nonlinear downscaling models and compared with linear MLR downscaling model. In the final step (step 3) of this study, future projection of precipitation and temperature was conducted. Therefore, the results in this study were presented accordingly.

Step 1 results (selection of downscaling predictors for predictands)

The results of the applied MI feature extraction technique are presented in Tables 3 and 4. It should be understood that a predictor with higher MI signifies the dominant predictor for the predictand.

Table 3

Step 1 results across the study stations for MI input screening

StationPredictorPrecipitation
Temperature
MIRankMIRank
Famagusta Tasmax 1.6374 1.5568 
Ts 1.6293 1.5403 
Hurs 1.6460 1.5687 
Huss 1.6735 1.5972 
Psl 1.6604 1.5480 
Evspsbl 1.6069 1.5253 10 
Sfcwind 1.6633 1.5850 
Ps 1.6297 1.5272 
Uas 1.6698 1.5968 
Vas 1.6015 10 1.5188 11 
Prw 1.4028 12 1.3135 13 
Pr 1.5086 11 1.4112 12 
Clwvi 1.3674 13 1.6035 
Nicosia Tasmax 1.6335 1.5868 
Ts 1.6261 1.5645 
Hurs 1.6783 1.6059 
Huss 1.7020 1.6316 
Psl 1.6732 1.6168 
Evspsbl 1.5368 10 1.4711 11 
Sfcwind 1.6776 1.6213 
Ps 1.6071 1.5476 
Uas 1.6951 1.6189 
Vas 1.5866 1.5342 10 
Prw 1.3786 13 1.3169 13 
Pr 1.4906 11 1.4218 12 
Clwvi 1.3891 12 1.6311 
Kyrenia Tasmax 1.6902 1.6161 
Ts 1.7483 1.6622 
Hurs 1.7367 1.6362 
Huss 1.7468 1.6484 
Psl 1.7621 1.6567 
Evspsbl 1.7393 1.6289 
Sfcwind 1.7707 1.6492 
Ps 1.7149 1.6031 10 
Uas 1.7727 1.6648 
Vas 1.6661 10 1.5503 11 
Prw 1.4755 12 1.3610 13 
Pr 1.5585 11 1.4506 12 
Clwvi 1.4386 13 1.6566 
StationPredictorPrecipitation
Temperature
MIRankMIRank
Famagusta Tasmax 1.6374 1.5568 
Ts 1.6293 1.5403 
Hurs 1.6460 1.5687 
Huss 1.6735 1.5972 
Psl 1.6604 1.5480 
Evspsbl 1.6069 1.5253 10 
Sfcwind 1.6633 1.5850 
Ps 1.6297 1.5272 
Uas 1.6698 1.5968 
Vas 1.6015 10 1.5188 11 
Prw 1.4028 12 1.3135 13 
Pr 1.5086 11 1.4112 12 
Clwvi 1.3674 13 1.6035 
Nicosia Tasmax 1.6335 1.5868 
Ts 1.6261 1.5645 
Hurs 1.6783 1.6059 
Huss 1.7020 1.6316 
Psl 1.6732 1.6168 
Evspsbl 1.5368 10 1.4711 11 
Sfcwind 1.6776 1.6213 
Ps 1.6071 1.5476 
Uas 1.6951 1.6189 
Vas 1.5866 1.5342 10 
Prw 1.3786 13 1.3169 13 
Pr 1.4906 11 1.4218 12 
Clwvi 1.3891 12 1.6311 
Kyrenia Tasmax 1.6902 1.6161 
Ts 1.7483 1.6622 
Hurs 1.7367 1.6362 
Huss 1.7468 1.6484 
Psl 1.7621 1.6567 
Evspsbl 1.7393 1.6289 
Sfcwind 1.7707 1.6492 
Ps 1.7149 1.6031 10 
Uas 1.7727 1.6648 
Vas 1.6661 10 1.5503 11 
Prw 1.4755 12 1.3610 13 
Pr 1.5585 11 1.4506 12 
Clwvi 1.4386 13 1.6566 
Table 4

Step 1 results across the study stations for CC input screening

StationPredictorPrecipitation
Temperature
CCRankCCRank
Famagusta Tasmax −0.6220 13 0.9686 
Ts −0.4911 10 0.9379 
Hurs −0.2787 0.1731 
Huss −0.5641 12 0.9459 
Psl 0.6356 −0.8804 13 
Evspsbl 0.5168 −0.4202 
Sfcwind −0.3216 0.5200 
Ps 0.5473 −0.6790 
Uas −0.3523 0.2119 
Vas 0.5482 −0.8714 12 
Prw −0.5378 11 0.8505 
Pr 0.5162 −0.7246 10 
Clwvi 0.5781 −0.7493 11 
Nicosia Tasmax −0.6230 13 0.9677 
Ts −0.4956 11 0.9304 
Hurs −0.3636 0.5010 
Huss −0.5689 12 0.9611 
Psl 0.5832 −0.7114 11 
Evspsbl 0.2197 0.0349 
Sfcwind −0.0862 0.0042 
Ps 0.5168 −0.6915 
Uas −0.1020 0.0838 
Vas 0.5563 −0.8640 13 
Prw −0.4651 10 0.8102 
Pr 0.4995 −0.7094 10 
Clwvi 0.5750 −0.7584 12 
Kyrenia Tasmax −0.6184 13 0.9732 
Ts −0.5153 11 0.9400 
Hurs 0.6448 −0.8769 13 
Huss −0.5814 12 0.9551 
Psl 0.6209 −0.7112 
Evspsbl 0.1475 0.0296 
Sfcwind −0.0963 0.0871 
Ps 0.6233 −0.8763 12 
Uas −0.2456 0.0501 
Vas 0.5004 −0.8757 11 
Prw −0.4651 10 0.7664 
Pr 0.5295 −0.7217 
Clwvi 0.6184 −0.7636 10 
StationPredictorPrecipitation
Temperature
CCRankCCRank
Famagusta Tasmax −0.6220 13 0.9686 
Ts −0.4911 10 0.9379 
Hurs −0.2787 0.1731 
Huss −0.5641 12 0.9459 
Psl 0.6356 −0.8804 13 
Evspsbl 0.5168 −0.4202 
Sfcwind −0.3216 0.5200 
Ps 0.5473 −0.6790 
Uas −0.3523 0.2119 
Vas 0.5482 −0.8714 12 
Prw −0.5378 11 0.8505 
Pr 0.5162 −0.7246 10 
Clwvi 0.5781 −0.7493 11 
Nicosia Tasmax −0.6230 13 0.9677 
Ts −0.4956 11 0.9304 
Hurs −0.3636 0.5010 
Huss −0.5689 12 0.9611 
Psl 0.5832 −0.7114 11 
Evspsbl 0.2197 0.0349 
Sfcwind −0.0862 0.0042 
Ps 0.5168 −0.6915 
Uas −0.1020 0.0838 
Vas 0.5563 −0.8640 13 
Prw −0.4651 10 0.8102 
Pr 0.4995 −0.7094 10 
Clwvi 0.5750 −0.7584 12 
Kyrenia Tasmax −0.6184 13 0.9732 
Ts −0.5153 11 0.9400 
Hurs 0.6448 −0.8769 13 
Huss −0.5814 12 0.9551 
Psl 0.6209 −0.7112 
Evspsbl 0.1475 0.0296 
Sfcwind −0.0963 0.0871 
Ps 0.6233 −0.8763 12 
Uas −0.2456 0.0501 
Vas 0.5004 −0.8757 11 
Prw −0.4651 10 0.7664 
Pr 0.5295 −0.7217 
Clwvi 0.6184 −0.7636 10 

As depicted by the MI results in Table 3, different predictors characterize different performance for certain predictants for different stations. For Famagusta station, the results show that Huss (1.6735), Uas (1.6698), and Sfcwind (1.6633) are, respectively, ranked as the first, second, and third most dominant predictors with respect to precipitation predictand. These are relative humidity and wind speed related predictors. According to the results provided by Allen et al. (1998) study, relative humidity and wind speed are important parameters in the hydrologic water cycle whereas precipitation is among the most significant parameters. This shows that the relationship and bonding (between precipitation, relative humidity, and wind speed) are not affected by local-scale climate but they are strong for large-scale climate. For temperature, the most dominant predictors are Clwvi (1.6035), Huss (1.5972), and Uas. This shows that both Huss and Uas have a strong nonlinear relationship with the predictands (precipitation and temperature). This may be because Famagusta is a semi-arid coastal station which is both affected by the characteristics of hot temperature and insufficiency of rainfall.

Despite the nonlinearity effect displayed by predictors, it is a well-known fact that time series data may contain both linear and nonlinear aspects not only for local-scale data but also for large-scale data. Therefore, in order to determine the linear relationship between the predictors and the predictands, the CC method of input screening was performed. The results of the applied CC method of feature extraction are presented in Table 4.

In the results shown in Table 4, it can be seen that some variables have a negative relationship with the predictands while some have positive values. For Famagusta station, considering precipitation as the predictand, the pressure-related predictor (Psl) was found to be most effective followed by wind speed-related predictors (Clwvi and Vas). By comparing with the results achieved by the MI method, in Famagusta station, it is obvious that due to the linear and nonlinear nature of the feature extraction methods, distinct results were obtained. For example, Clwvi is found to be the least effective predictor variable using the MI method, while the CC method showed Clwvi as the most dominant predictor. However, as shown in Table 4, Tasmax, Huss, and Ts are, respectively, the best, second, and third most sensitive predictors with temperature as the predictand. This demonstrates that the linear correlation is stronger towards temperature variable. With the determination of the significant predictors, the statistical downscaling modeling was then performed and the results are presented in the next section.

Step 2 results (statistical downscaling modeling)

The network parameters used to develop the ANN model including learning algorithm, transfer function type, training function, etc. for the two predictands under study for all stations are presented in Table 5.

Table 5

ANN network parameters for statistical downscaling modeling

StationModel functionPrecipitation
Temperature
MICCMICC
Famagusta Type of network used FFNN FFNN FFNN FFNN 
Learning algorithm LM LM LM LM 
Training function TrainLM TrainLM TrainLM TrainLM 
Number of input layer neurons 
Number of hidden layer neurons 12 10 
Number of output layer neurons 
Transfer function type TANSIG TANSIG TANSIG TANSIG 
Maximum epochs 1,000 1,000 1,000 1,000 
Number of iterations 236 198 70 150 
Nicosia Type of network used FFNN FFNN FFNN FFNN 
Learning algorithm LM LM LM LM 
Training function TrainLM TrainLM TrainLM TrainLM 
Number of input layer neurons 
Number of hidden layer neurons 
Number of output layer neurons 
Transfer function type TANSIG TANSIG TANSIG TANSIG 
Maximum epochs 1,000 1,000 1,000 1,000 
Number of iterations 300 276 300 276 
Kyrenia Type of network used FFNN FFNN FFNN FFNN 
Learning algorithm LM LM LM LM 
Training function TrainLM TrainLM TrainLM TrainLM 
Number of input layer neurons 
Number of hidden layer neurons 10 
Number of output layer neurons 
Transfer function type TANSIG TANSIG TANSIG TANSIG 
Maximum epochs 1,000 1,000 1,000 1,000 
Number of iterations 120 96 120 96 
StationModel functionPrecipitation
Temperature
MICCMICC
Famagusta Type of network used FFNN FFNN FFNN FFNN 
Learning algorithm LM LM LM LM 
Training function TrainLM TrainLM TrainLM TrainLM 
Number of input layer neurons 
Number of hidden layer neurons 12 10 
Number of output layer neurons 
Transfer function type TANSIG TANSIG TANSIG TANSIG 
Maximum epochs 1,000 1,000 1,000 1,000 
Number of iterations 236 198 70 150 
Nicosia Type of network used FFNN FFNN FFNN FFNN 
Learning algorithm LM LM LM LM 
Training function TrainLM TrainLM TrainLM TrainLM 
Number of input layer neurons 
Number of hidden layer neurons 
Number of output layer neurons 
Transfer function type TANSIG TANSIG TANSIG TANSIG 
Maximum epochs 1,000 1,000 1,000 1,000 
Number of iterations 300 276 300 276 
Kyrenia Type of network used FFNN FFNN FFNN FFNN 
Learning algorithm LM LM LM LM 
Training function TrainLM TrainLM TrainLM TrainLM 
Number of input layer neurons 
Number of hidden layer neurons 10 
Number of output layer neurons 
Transfer function type TANSIG TANSIG TANSIG TANSIG 
Maximum epochs 1,000 1,000 1,000 1,000 
Number of iterations 120 96 120 96 

FFNN, LM, and TANSIG represent feed forward neural network, Levenberg–Marquardt, and tangent sigmoid, respectively.

For ANFIS downscaling modeling, hybrid optimization algorithm was used for the calibration of Sugeno type fuzzy inference system. Several membership functions were used including Trapezoidal, Triangular and Gaussian. To achieve the best ANFIS construction, trial and error procedure was employed for the network selection. The parameters used in developing ANFIS downscaling modeling are presented in Table 6.

Table 6

ANFIS network parameters for statistical downscaling modeling

StationModel functionPrecipitation
Temperature
MICCMICC
Famagusta Learning algorithm Hybrid Hybrid Hybrid Hybrid 
Fuzzy inference algorithm Sugeno Sugeno Sugeno Sugeno 
Structure formulation TA TA TA TA 
Membership function type Triangular Trapezoidal Gaussian Triangular 
Generated FIS GP GP GP GP 
Number of membership functions 
Output type Constant Constant Constant Constant 
Error tolerance 0.0005 0.0005 0.0005 0.0005 
Epochs 53 29 60 77 
Nicosia Learning algorithm Hybrid Hybrid Hybrid Hybrid 
Fuzzy inference algorithm Sugeno Sugeno Sugeno Sugeno 
Structure formulation TA TA TA TA 
Membership function type Trapezoidal Trapezoidal Gaussian Gaussian 
Generated FIS GP GP GP GP 
Number of membership functions 
Output type Constant Constant Constant Constant 
Error tolerance 0.0001 0.0001 0.0001 0.0003 
Epochs 70 102 39 82 
Kyrenia Learning algorithm Hybrid Hybrid Hybrid Hybrid 
Fuzzy inference algorithm Sugeno Sugeno Sugeno Sugeno 
Structure formulation TA TA TA TA 
Membership function type Triangular Gaussian Gaussian Trapezoidal 
Generated FIS GP GP GP GP 
Number of membership functions 
Output type Constant Constant Constant Constant 
Error tolerance 0.0005 0.0005 0.0005 0.0005 
Epochs 88 110 90 68 
StationModel functionPrecipitation
Temperature
MICCMICC
Famagusta Learning algorithm Hybrid Hybrid Hybrid Hybrid 
Fuzzy inference algorithm Sugeno Sugeno Sugeno Sugeno 
Structure formulation TA TA TA TA 
Membership function type Triangular Trapezoidal Gaussian Triangular 
Generated FIS GP GP GP GP 
Number of membership functions 
Output type Constant Constant Constant Constant 
Error tolerance 0.0005 0.0005 0.0005 0.0005 
Epochs 53 29 60 77 
Nicosia Learning algorithm Hybrid Hybrid Hybrid Hybrid 
Fuzzy inference algorithm Sugeno Sugeno Sugeno Sugeno 
Structure formulation TA TA TA TA 
Membership function type Trapezoidal Trapezoidal Gaussian Gaussian 
Generated FIS GP GP GP GP 
Number of membership functions 
Output type Constant Constant Constant Constant 
Error tolerance 0.0001 0.0001 0.0001 0.0003 
Epochs 70 102 39 82 
Kyrenia Learning algorithm Hybrid Hybrid Hybrid Hybrid 
Fuzzy inference algorithm Sugeno Sugeno Sugeno Sugeno 
Structure formulation TA TA TA TA 
Membership function type Triangular Gaussian Gaussian Trapezoidal 
Generated FIS GP GP GP GP 
Number of membership functions 
Output type Constant Constant Constant Constant 
Error tolerance 0.0005 0.0005 0.0005 0.0005 
Epochs 88 110 90 68 

GP and TA represent grid partitioning and trial and error.

It is important to state that since the study is concerned with statistical downscaling modeling of precipitation and temperature, the results are presented in different subsections for all stations.

Statistical downscaling of precipitation

The results of the downscaling modeling of precipitation using the ANN, ANFIS, and MLR downscaling models for both MI and CC predictor screening methods across the study stations are provided in Table 7. It is worth clarifying that the results are presented for the best downscaling performance.

Table 7

Statistical downscaling results for precipitation

StationInput screening methodDownscaling modelInputsTraining
Validation
DCaRMSEDCaRMSE
Famagusta MI ANN Huss, Uas, Sfcwind 0.589 0.097 0.554 0.121 
ANFIS Huss, Uas, Sfcwind 0.546 0.101 0.518 0.126 
MLR Huss, Uas, Sfcwind 0.404 0.116 0.377 0.143 
CC ANN Psl, Clwvi, Vas 0.438 0.136 0.360 0.120 
ANFIS Psl, Clwvi, Vas 0.507 0.106 0.497 0.129 
MLR Psl, Clwvi, Vas 0.433 0.137 0.355 0.121 
Nicosia MI ANN Huss, Uas, Hurs 0.408 0.116 0.404 0.142 
ANFIS Huss, Uas, Hurs 0.450 0.127 0.286 0.137 
MLR Huss, Uas, Hurs 0.337 0.150 0.291 0.126 
CC ANN Psl, Clwvi, Vas 0.570 0.121 0.404 0.116 
ANFIS Psl, Clwvi, Vas 0.563 0.122 0.394 0.117 
MLR Psl, Clwvi, Vas 0.457 0.136 0.368 0.119 
Kyrenia MI ANN Uas, Sfcwind, Psl 0.556 0.099 0.487 0.130 
ANFIS Uas, Sfcwind, Psl 0.532 0.102 0.581 0.118 
MLR Uas, Sfcwind, Psl 0.435 0.137 0.413 0.114 
CC ANN Hurs, Ps, Psl 0.522 0.126 0.371 0.118 
ANFIS Hurs, Ps, Psl 0.562 0.120 0.492 0.106 
MLR Hurs, Ps, Psl 0.475 0.132 0.469 0.108 
StationInput screening methodDownscaling modelInputsTraining
Validation
DCaRMSEDCaRMSE
Famagusta MI ANN Huss, Uas, Sfcwind 0.589 0.097 0.554 0.121 
ANFIS Huss, Uas, Sfcwind 0.546 0.101 0.518 0.126 
MLR Huss, Uas, Sfcwind 0.404 0.116 0.377 0.143 
CC ANN Psl, Clwvi, Vas 0.438 0.136 0.360 0.120 
ANFIS Psl, Clwvi, Vas 0.507 0.106 0.497 0.129 
MLR Psl, Clwvi, Vas 0.433 0.137 0.355 0.121 
Nicosia MI ANN Huss, Uas, Hurs 0.408 0.116 0.404 0.142 
ANFIS Huss, Uas, Hurs 0.450 0.127 0.286 0.137 
MLR Huss, Uas, Hurs 0.337 0.150 0.291 0.126 
CC ANN Psl, Clwvi, Vas 0.570 0.121 0.404 0.116 
ANFIS Psl, Clwvi, Vas 0.563 0.122 0.394 0.117 
MLR Psl, Clwvi, Vas 0.457 0.136 0.368 0.119 
Kyrenia MI ANN Uas, Sfcwind, Psl 0.556 0.099 0.487 0.130 
ANFIS Uas, Sfcwind, Psl 0.532 0.102 0.581 0.118 
MLR Uas, Sfcwind, Psl 0.435 0.137 0.413 0.114 
CC ANN Hurs, Ps, Psl 0.522 0.126 0.371 0.118 
ANFIS Hurs, Ps, Psl 0.562 0.120 0.492 0.106 
MLR Hurs, Ps, Psl 0.475 0.132 0.469 0.108 

aRMSE: Being normalized data, RMSE is unitless.

As given in Table 7, the downscaling performances of the models are evaluated according to DC and RMSE performance indicators. Considering Famagusta station, it can be seen that the AI downscaling models performed better than MLR downscaling model with respect to both MI and CC predictors' screening methods. In terms of DC and based on validation results, ANN is found to be the more efficient downscaling model with 4 and 17% over ANFIS and MLR, respectively, based on MI feature extraction. In terms of CC, ANFIS outperformed the other models with 14 and 15% over ANN and MLR, respectively.

For Nicosia station, the general performance of the downscaling models is lower in comparison to Famagusta station, especially that obtained by MI method. This could be because of the difference in location of the two stations (inland for Nicosia and coastal for Famagusta). The better performance of models produced by the CC method could be due to the direct effect the environment has on precipitation aspect, in the sense that the energy reaching the area causes a direct increase in temperature and is easily predictable with less fluctuations which result in an almost constant amount of evaporation and transpiration and hence precipitation.

For Kyrenia station, it is expected that the performance of the statistical downscaling models would be higher than in Famagusta and Nicosia stations. This is as a result of a higher amount and frequency of precipitation experienced in Kyrenia station. As the results in Table 7 show, a high amount of precipitation received in an area signifies better downscaling capabilities and, of course, better performance and vice versa. Similar to the coastal station (Famagusta), ANFIS is found to be more efficient than the rest of the downscaling models. To evaluate the monthly precipitation downscaling performance of the applied models, the observed monthly precipitation sum is compared with the downscaled precipitation sum by the applied downscaling models and presented in Figure 5.

Figure 5

Comparison of monthly precipitation between observed and downscaling models for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

Figure 5

Comparison of monthly precipitation between observed and downscaling models for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

It can be seen from Figure 5 that December and January are the months with the highest precipitation amount and the months span from June to August are periods with little or no rainfall amounts across all the study stations. In view of the closeness and fitness of the downscaled models towards the observed precipitation between June and August it can be deduced that the downscaling models are more reliable within that period.

Statistical downscaling of mean temperature

Similarly to precipitation, the results of the downscaling modeling of temperature using the ANN, ANFIS, and MLR downscaling models for both MI and CC predictors' screening methods across the study stations are provided in Table 8. However, the results are presented for the best downscaling performance.

Table 8

Statistical downscaling results for temperature

StationInput screening methodDownscaling modelInputsTraining
Validation
DCaRMSEDCaRMSE
Famagusta MI ANN Clwvi, Huss, Uas 0.948 0.069 0.940 0.077 
ANFIS Clwvi, Huss, Uas 0.951 0.067 0.943 0.075 
MLR Clwvi, Huss, Uas 0.923 0.085 0.914 0.092 
CC ANN Tasmax, Huss, Ts 0.976 0.047 0.976 0.049 
ANFIS Tasmax, Huss, Ts 0.972 0.052 0.969 0.053 
MLR Tasmax, Huss, Ts 0.954 0.067 0.951 0.067 
Nicosia MI ANN Huss, Clwvi, Sfcwind 0.945 0.071 0.944 0.073 
ANFIS Huss, Clwvi, Sfcwind 0.965 0.056 0.959 0.063 
MLR Huss, Clwvi, Sfcwind 0.950 0.067 0.941 0.076 
CC ANN Tasmax, Huss, Ts 0.970 0.052 0.970 0.054 
ANFIS Tasmax, Huss, Ts 0.968 0.054 0.968 0.056 
MLR Tasmax, Huss, Ts 0.953 0.065 0.951 0.069 
Kyrenia MI ANN Uas, Ts, Psl 0.958 0.062 0.954 0.066 
ANFIS Uas, Ts, Psl 0.962 0.058 0.959 0.063 
MLR Uas, Ts, Psl 0.927 0.081 0.926 0.084 
CC ANN Tasmax, Huss, Ts 0.966 0.058 0.963 0.058 
ANFIS Tasmax, Huss, Ts 0.970 0.053 0.967 0.055 
MLR Tasmax, Huss, Ts 0.953 0.067 0.953 0.065 
StationInput screening methodDownscaling modelInputsTraining
Validation
DCaRMSEDCaRMSE
Famagusta MI ANN Clwvi, Huss, Uas 0.948 0.069 0.940 0.077 
ANFIS Clwvi, Huss, Uas 0.951 0.067 0.943 0.075 
MLR Clwvi, Huss, Uas 0.923 0.085 0.914 0.092 
CC ANN Tasmax, Huss, Ts 0.976 0.047 0.976 0.049 
ANFIS Tasmax, Huss, Ts 0.972 0.052 0.969 0.053 
MLR Tasmax, Huss, Ts 0.954 0.067 0.951 0.067 
Nicosia MI ANN Huss, Clwvi, Sfcwind 0.945 0.071 0.944 0.073 
ANFIS Huss, Clwvi, Sfcwind 0.965 0.056 0.959 0.063 
MLR Huss, Clwvi, Sfcwind 0.950 0.067 0.941 0.076 
CC ANN Tasmax, Huss, Ts 0.970 0.052 0.970 0.054 
ANFIS Tasmax, Huss, Ts 0.968 0.054 0.968 0.056 
MLR Tasmax, Huss, Ts 0.953 0.065 0.951 0.069 
Kyrenia MI ANN Uas, Ts, Psl 0.958 0.062 0.954 0.066 
ANFIS Uas, Ts, Psl 0.962 0.058 0.959 0.063 
MLR Uas, Ts, Psl 0.927 0.081 0.926 0.084 
CC ANN Tasmax, Huss, Ts 0.966 0.058 0.963 0.058 
ANFIS Tasmax, Huss, Ts 0.970 0.053 0.967 0.055 
MLR Tasmax, Huss, Ts 0.953 0.067 0.953 0.065 

aRMSE: Being normalized data, RMSE is unitless.

By visual inspection of Tables 7 and 8 it can be understood that the general performance of the downscaling models are more superior in temperature downscaling than precipitation downscaling. This may be because of the distinct characteristics of the predictands in the study stations. Having a semi-arid Mediterranean climate, the stations under study are subjected to high temperature and scarce water resources (less or little rainfall). The harsh weather, coupled with uncertainty in factors affecting hydrologic cycle, leads to a greater stochastic nature of precipitation and, hence, difficulty in downscaling. The results in Table 8 show better downscaling capability of AI models than MLR model across all stations.

For Famagusta station in the validation phase based on DC performance indicator, the results in Table 8 revealed that ANFIS produced the best performance in terms of MI. ANFIS improved performance in the mean temperature downscaling modeling up to 1 and 3% over ANN and MLR, respectively. Regarding the CC feature screening method, ANN was the best performing downscaling model with an improvement of 1% over ANFIS and 3% over MLR. For Nicosia station, ANFIS had the best performance with DC 0.959, followed by ANN with 0.944 and, finally, MLR with 0.941 with respect to MI, whereas ANN slightly edged out ANFIS with DCs of 0.970 and 0.968, and MLR followed with DC of 0.951. For Kyrenia station, ANFIS was narrowly the best downscaling model with DC of 0.959 and RMSE 0.063, followed by ANN with DC of 0.954 and RMSE 0.066. The MLR downscaling model performed the least with DC of 0.926 and RMSE 0.084. To ascertain the performances of the downscaling models for mean temperature in the period of this study (1995–2017), the monthly average of the statistically downscaled temperature for each month is compared with the observed mean temperature, as shown in Figure 6.

Figure 6

Comparison of monthly mean temperature between observed and downscaling models for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

Figure 6

Comparison of monthly mean temperature between observed and downscaling models for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

Figure 6 shows the downscaling models are very close to the observed values which ultimately confirmed the performance illustrated by DC and RMSE efficiency criteria. However, when the downscaling results of Figures 5 and 6 are compared, it can be deduced that values in Figure 6 are more fitted to observed values. This indicates that better downscaling modeling is achieved with temperature as predictand than with precipitation. Moreover, Figure 6 indicates that temperature within the study region is higher in hot months (most especially the hottest months of June, July, August, and September) and lower in cold months (December, January, February, and March).

Step 3 results (future projection of precipitation and temperature)

The results for the future projection of precipitation and temperature are provided in two subsections.

Results of precipitation projection for the future

It can be seen from Table 8 that in most cases ANN provided the best performance for the downscaling modeling using MI feature extraction method, while ANFIS mostly performed better for CC method. Therefore, for future projection of precipitation, ANN was used as the benchmark model for projection made based on MI method and ANFIS was used as the benchmark model for projection based on CC method. Figure 7 compares the benchmark monthly precipitation sum and future projection by the applied models.

Figure 7

Monthly projected precipitation sum in the future (2018–2040) for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

Figure 7

Monthly projected precipitation sum in the future (2018–2040) for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

Since the projection results showed that the ANN model performed better in the projection of precipitation for the future, the discussion in the section is done through comparison between benchmark precipitation and future projection by ANN model only. Considering the projected precipitation sum for Famagusta station it can be seen that, in almost every month, there would be a decrease in the precipitation amount with the exception of May and August. This could be because the two months serve as the strategic points for the arrival of summer and the entry to winter. It can be seen that periods with higher precipitation including November, December, January, February, and March will experience a greater decrease in precipitation than months with less frequency and amount of precipitation. For instance, a decrease of about 8% could be noticed in November, 14% in December, 12% in January and February, and 5% in March. The results in Figure 7 show a different pattern of precipitation in Nicosia with respect to Famagusta. The precipitation trend shows an increase in precipitation in the summer period (March, May, June, July, August, September) and decreasing trend in October, November, December, January, and February while April remains the same. However, the precipitation decrease is very low in Nicosia compared to Famagusta. For instance, a decrease of 0.5, 1, 2, 2, and 2% is observed for October, November, December, January, and February whereas an increase in precipitation values of up to 0.5, 4, 4, 0.2, 2, and 1% are witnessed for March, May, June, July, August, and September, respectively. For Kyrenia station, despite having a different amount of precipitation sum compared to Famagusta, it has a similar percentage decrease in precipitation. The only difference is that throughout the period, there is no increasing trend for precipitation in any month and there is a large percentage decreasing every year.

The general results for the future projection of precipitation showed that towards the middle of the 21st century (2018–2040), the amount of precipitation is expected to decrease within the stations under consideration. Kyrenia station is expected to be more impacted with a reduction of up to 42%. This is because the station experiences the highest precipitation in North Cyprus and it is obvious that the higher precipitation received by a region, the more availability of water to be lost by evapotranspiration and, thus, the difference between the inflows and outflows will be larger compared with an area that received less precipitation. This is proven by the results of this study. However, Famagusta station, due to the presence of sea, receives the second highest precipitation and, hence, second highest decrease up to 22% of precipitation towards 2040. On the other hand, Nicosia being an inland station receives less precipitation and, subsequently, less precipitation decrease of 4.8%.

Results of temperature projection for the future

As depicted by Table 8, both AI downscaling models performed efficiently and, at one point in a certain station and certain input screening method, ANN provided the most reliable performance, whereas in another circumstance, ANFIS is most significant. Moreover, their difference in performance is very small, to the extent that any of them could be efficient if selected for the future projection of temperature. In view of this, ANN statistical downscaled values were selected as the benchmark temperature for all stations. The performance of the models for the future projection of temperature in most cases is higher in Nicosia station for both MI and CC input selection methods. This could be because of the slightly higher temperature that Nicosia is experiencing in comparison to Famagusta and Kyrenia stations (see Table 3). Comparing the performances across the stations, it can be seen that for Famagusta, ANN provided the best performance based on MI, whereas for CC method, ANFIS led to better performance. Since all the three models have shown accuracy in the future projection of temperature, the temperature values produced by the models are compared with the benchmark temperature and presented in Figure 8 to determine the trend of decrease or increase in temperature between 2018 and 2040.

Figure 8

Monthly projected mean temperature in the future (2018–2040) for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

Figure 8

Monthly projected mean temperature in the future (2018–2040) for (a) Famagusta station, (b) Nicosia station, and (c) Kyrenia station.

The close temperature trend shown by the models in Figure 8 affirmed the reliable performance of the models indicated by DC and RMSE evaluation criteria in Table 8. However, the future projection of higher mean temperature produced by the MLR model shown in Figure 8 could be due to error produced by the model as can be seen in Table 8. Since ANN and ANFIS models are found to be more robust models as shown in both Table 8 and Figure 8, the discussion regarding the mean temperature for the future is done based on the two models.

For Famagusta station, both ANN and ANFIS models show an increase in the future temperature of less than 1% in the months of January and February. March, May, June, September, November, and December show no difference (decrease or increase) in the mean temperature by ANN model but for the ANFIS model, the months with no difference are April and August only. A decrease in the temperature is noticed in April, July, August, and October. For Nicosia station, an increase in the mean temperature will occur from January through June. This may be because with continuing greenhouse effect, the temperature of hot weather climate increases and gradually even the temperature in cold weather changes. A short decrease of usually less than 1% of mean temperature will occur from July to September while October, November, and December will also witness a rise in temperature. What this implies is that the future temperature for Nicosia will increase in the winter season and slightly increase in summer season. For Kyrenia station, the increase and decrease in the mean temperature is similar to Nicosia station. From November through March the temperature is expected to rise. While April and May show a balance between previous and future mean temperature, June, July, August, September, and October will experience a slight decrease.

The general results of the projected precipitation and mean temperature showed that there will be a decrease in precipitation amount in Kyrenia station in every month of the year from 2018 to 2040, while a decrease in precipitation is also expected for Nicosia and Famagusta stations for all months with the exception of May where the results showed an expected increase in precipitation. With regards to mean temperature, both increase and decrease are possible due to the effect of greenhouse gases. The increase will occur mostly in winter periods while summer periods will experience a decrease in temperature. The results also showed that in terms of percentage, a higher percentage of precipitation would be lost compared to an increment or decrease percentage of mean temperature.

In this paper, the impact of climate change on hydro-climatological parameters, especially precipitation and temperature, in North Cyprus was evaluated for a period from 2018 to 2040. At a grid point located in Karfas, 13 predictors from BNU-ESM GCM based on CMIP5 under RCP4.5 were used for statistical downscaling and future projection of mean precipitation and temperature over Famagusta, Nicosia, and Kyrenia stations using ANN, ANFIS, and MLR models. MI and CC feature extraction methods were used to determine both nonlinear and linear relationship between the predictors/predictands and, later, the most dominant predictors were used as input variables for the downscaling modeling. The results of the best downscaling model were then used as a benchmark for the future projection of the hydro-climatological parameters.

The results of the feature extraction methods showed that for both precipitation and temperature, different predictors led to different performance. For Famagusta station, based on precipitation, a combination of Huss, Uas, Sfcwind for MI and Psl, Clwvi, Vas for CC provided the best performance, while Clwvi, Huss, Uas for MI and Tasmax, Huss, Ts led to better accuracy based on temperature. For Nicosia station, based on precipitation, a combination of Huss, Uas, Hurs for MI and Psl, Clwvi, Vas for CC provided the best performance, while Huss, Clwvi, Sfcwind for MI and Tasmax, Huss, Ts led to better accuracy based on temperature. For Kyreniastation, based on precipitation, a combination of Uas, Sfcwind, Psl for MI and Hurs, Ps, Psl for CC provided the best performance, while Uas, Ts, Psl for MI and Tasmax, Huss, Ts led to better accuracy based on temperature. The results of the downscaling modeling showed that the AI-based models performed better than the MLR model. For instance, in terms of precipitation, ANFIS outperformed the other models with 14 and 15% over ANN and MLR, respectively. In terms of temperature, more reliable downscaling was achieved by ANFIS, which has 1.5% and 1.8% higher accuracy than ANN and MLR, respectively. The future projection of the predictands showed a decreasing trend in the annual precipitation and an increasing trend in temperature. This suggests that the greenhouse effect will continue to escalate in the years to come and thus increase global warming with a precipitation decrease of up to 22%, 4.8%, 42% and temperature increase up to 2.9%, 5%, 4.8% for Famagusta, Nicosia, and Kyrenia stations, respectively.

The statistical downscaling models showed better skill for temperature than precipitation. The ensemble model that combines other models shows superior performance in statistical downscaling modeling, such as in the case of Nourani et al. (2018) study, therefore, for future studies, the ensemble model should be employed to improve the downscaling performance of precipitation. However, this study employed CMIP5 data to perform the statistical downscaling; with the emergence of CMIP6 data, perhaps more rigorous results could be achieved. It is therefore suggested to use CMIP6 data for future studies and compare.

Not applicable.

The research has no conflict of interest.

The codes are available in the modules of software used.

All relevant data are included in the paper or its Supplementary Information.

Allen
R. G.
,
Pereira
L. S.
,
Raes
D.
&
Smith
M.
1998
Crop Evapotranspiration-Guidelines for Computing Crop Water Requirements
.
FAO Irrigation and drainage paper 56
.
FAO
,
Rome, Italy, D05109
.
Baghanam
A. H.
,
Nourani
V.
,
Keynejad
M. A.
,
Taghipour
H.
&
Alami
M. T.
2019
Conjunction of wavelet-entropy and SOM clustering for multi-GCM statistical downscaling
.
Hydrology Research
50
(
1
),
1
23
.
Bowden
G. J.
,
Dandy
G. C.
&
Maier
H. R.
2005
Input determination for neural network models in water resources applications. Part 1—background and methodology
.
Journal of Hydrology
301
(
1–4
),
75
92
.
Hornik
K.
,
Stinchcombe
M.
&
White
H.
1989
Multilayer feedforward networks are universal approximators
.
Neural Networks
2
(
5
),
359
366
.
Jang
J. S.
1993
ANFIS: Adaptive-Network-Based Fuzzy Inference System
.
IEEE Transactions on Systems, Man, and Cybernetics
23
,
665
685
.
Jang
J. R.
&
Sun
C. T.
1995
Neuro-fuzzy modeling and control
.
Proceedings of the IEEE
83
(
3
),
378
406
.
Jang
J. S. R.
,
Sun
C. T.
&
Mizutani
E.
1997
Neuro-fuzzy and soft computing-a computational approach to learning and machine intelligence [Book review]
.
IEEE Transactions on Automatic Control
42
(
10
),
1482
1484
.
Landman
W. A.
,
Mason
S. J.
,
Tyson
P. D.
&
Tennant
W. J.
2001
Statistical downscaling of GCM simulations to streamflow
.
Journal of Hydrology
252
(
1–4
),
221
236
.
Mora
D. E.
,
Campozano
L.
,
Cisneros
F.
,
Wyseure
G.
&
Willems
P.
2014
Climate changes of hydrometeorological and hydrological extremes in the Paute basin
.
Ecuadorean Andes. Hydrology and Earth System Sciences
18
(
2
),
631
.
Nourani
V.
,
Khanghah
T. R.
&
Baghanam
A. H.
2015
Application of entropy concept for input selection of wavelet-ANN based rainfall-runoff modeling
.
Journal of Environmental Informatics
26
(
1
),
52
70
.
Nourani
V.
,
Razzaghzadeh
Z.
,
Baghanam
A. H.
&
Molajou
A.
2019a
ANN-based statistical downscaling of climatic parameters using decision tree predictor screening method
.
Theoretical and Applied Climatology
137
(
3–4
),
1729
1746
.
Nourani
V.
,
Elkiran
G.
&
Abdullahi
J.
2020
Multi-step ahead modeling of reference evapotranspiration using a multi-model approach
.
Journal of Hydrology
581
,
124434
.
Pahlavan
H. A.
,
Zahraie
B.
,
Nasseri
M.
&
Varnousfaderani
A. M.
2017
Improvement of multiple linear regression method for statistical downscaling of monthly precipitation
.
International Journal of Environmental Science and Technology
15
(
9
),
1897
1912
.
Peng
Y.
,
Zhao
X.
,
Wu
D.
,
Tang
B.
,
Xu
P.
,
Du
X.
&
Wang
H.
2018
Spatiotemporal variability in extreme precipitation in China from observations and projections
.
Water
10
(
8
),
1089
.
Samadi
S.
,
Carbone
G. J.
,
Mahdavi
M.
,
Sharifi
F.
&
Bihamta
M. R.
2013a
Statistical downscaling of river runoff in a semi-arid catchment
.
Water Resources Management
27
(
1
),
117
136
.
Samadi
S.
,
Wilson
C. A.
&
Moradkhani
H.
2013b
Uncertainty analysis of statistical downscaling models using hadley centre coupled model
.
Theoretical and Applied Climatology
114
(
3–4
),
673
690
.
Shannon
C. E.
1948
A mathematical theory of communication
.
The Bell System Technical |Journal
27
(
3
),
379
423
.
Timbal
B.
,
Dufour
A.
&
McAvaney
B.
2003
An estimate of future climate change for western France using a statistical downscaling technique
.
Climate Dynamics
20
(
7–8
),
807
823
.
Wilson
P.
&
Mantooth
H. A.
2013
Model-based Engineering for Complex Electronic Systems
.
Newnes
,
Oxford, UK
.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/).