Wavelet-based predictor screening for statistical downscaling of precipitation and temperature using the artificial neural network method

Throughout the last decades, industrial activities and the atmospheric concentration of greenhouse gases (GHGs) such as carbon dioxide (CO₂) have led to human-induced climate change. The resulting impacts of climate change including changes in the climate system and climate-related variables, such as precipitation and temperature, have become one of the major areas of concern for researchers. Therefore, examination of their future patterns will assist decision-makers to carry out effective measures to deal with the predicament such as droughts, flash floods, temperature extremes, and melting of ice shields, to name but a few. In this context, climate models have been developed, and they have the ability to predict the future variations in temperature and precipitation as a consequence of the increasing concentration of GHGs. Climate models employ physically based equations to simulate the processes governing the exchange of energy and mass between the atmosphere, lands, and oceans.

One of the main limitations of climate models is the coarse spatial resolutions of the outputs, inasmuch as many geographical features could be neglected or not represented in the models. To acquire the desirable local or regional scales of climate models data, several downscaling methods have been employed (Wilby & Wigley 1997). Downscaling approaches are divided into two categories. Dynamical downscaling is a method in which the outputs of climate models, as the boundary condition of regional climate models (RCMs), are applied to derive higher resolution. Demanding compute-intensive processes, dynamical downscaling methods are less economical and, therefore, challenging to apply. Statistical downscaling is a method to establish a statistical relationship between local-scale observed data (predictands) and large-scale climate variables (predictors). This is a more straightforward approach and can be run more rapidly, allowing researchers to simulate multiple models in a brief period. Furthermore, this method can be used for different regions (Timbal et al. 2003). Previous studies applied statistical downscaling methods such as statistical downscale models (SDSM; Wilby et al. 2002), K-nearest neighbors (Gangopadhyay et al. 2005), support vector machine (SVM; Tripathi et al. 2006), rule induction and tree algorithms (Goyal et al. 2012), and artificial neural network (ANN; Moghim & Bras 2017) to statistically downscale the climate variables. Over the last few years, machine learning methods such as ANN have been employed to investigate the nonlinear relationship between hydro-climatic variables as well as downscaling GCMs data. Studies conducted by Li & Zheng (2003), Dibike & Coulibaly (2006) evaluated the performance of ANN and regression models to downscale daily and monthly precipitation and temperature and concluded that the final results are similar. Another study conducted by Harpham & Wilby (2005) examined the performance of ANN and SDSM for downscaling the daily precipitation, and the results showed that ANN has the edge over SDSM. In a study by Tisseuil et al. (2010), the competitiveness of ANN and linear models such as the generalized linear model (GLM) was investigated. It was concluded that ANN performs better than linear models. Chadwick et al. (2011) proposed an improved ANN simulation to downscale temperature and rainfall. Su et al. (2015) used the ANN model to forecast the fluctuation of streamflow and precipitation during 2011–2050. Given the results of downscaling with machine learning techniques (e.g., ANN), there are some opposing conclusions. While some studies indicated the effectiveness of ANN application in downscaling GCMs data, others commented that ANN can lead to undesirable performance. Khan et al. (2006) in a comparative study analyzed the performance of ANN and SDSM, and the results demonstrated that SDSM performed more accurately in downscaling local climate features. In a study conducted by Abdellatif et al. (2013), the advantage of using the GLM-ANN model in downscaling was investigated and concluded that the proposed method has superiority over traditional ANN.

These contrary results appear to be rooted in the property and quantity of the datasets. It has been observed that large datasets can decrease the efficacy of ANN-based modeling and it would suffer from a data redundancy (Bowden et al. 2005). Then, it stands to reason that a data pre-processing technique should be applied to big datasets like the outputs of climate models to select the potential predictors among a large number of data (Nourani et al. 2017).

Among recent studies, Ahmadi & Han (2012) employed a synthesis method using Gamma test (GT) and coefficient correlation analysis to forecast the precipitation. Correlation analysis, as a common approach, was applied by Sachindra et al. (2013) to determine the dominant outputs of general circulation models (GCMs). Okkan (2015) utilized the mutual information (MI) and correlation method to select the potential predictors to downscale GCMs data. In a joint study, Devak & Dhanya (2016) investigated the performance of the principal component analysis (PCA) to decline the dimensionality of GCM variables. The correlation-based analysis is a conventional method to screen and derive the features of the large datasets. Nevertheless, one of the limitations of this method is its low ability to spot nonlinear relationships between time series. Therefore, in light of the fact that hydro-climatic variables have nonstationary and nonlinear features, traditional correlation methods would not lead to desirable results. To address this challenge, wavelet analysis has been properly employed by different studies to investigate nonstationary signals in hydro-climatological processes (Partal & Kişi 2007; Belayneh et al. 2014; Rashid et al. 2016; Tan et al. 2016; Xu et al. 2019).

In the field of statistical downscaling, Rana & Moradkhani (2016) employed continuous wavelet transform (CWT) in the projection step to analyze the possible variability in future precipitation and temperature. More recently, Baghanam et al. (2018) and Nourani et al. (2018a) utilized wavelet entropy (WE) to downscale the precipitation. Based on the preceding studies, CWT and the discrete wavelet transform (DWT) have been applied for feature extraction and denoising of a single signal, however, the application of the two methods has not been investigated for predictor screening purposes.

Cross-wavelet transform (XWT) is used as a technique to investigate and detect the coherence and phase relations among signals in hydro-climatic studies conducted by Adamowski (2008), Holman et al. (2011), Soon et al. (2014), and Yu & Lin (2015). Maraun & Kurths (2004) applied the XWT and wavelet coherence transform (WTC) to evaluate the relations between ENSO and NAO. Jevrejeva et al. (2003) described a WTC method to study the impacts of Arctic Oscillation (AO) and NAO on the state of the ice in the Baltic Sea. Grinsted et al. (2004) tested WTC to investigate the correlation between geophysical time series, namely the Arctic Oscillation (AO) and the Baltic maximum sea ice (BMI). Tamaddun et al. (2017) used CWT and wavelet correlation spectrum to examine the influence of oceanic-atmospheric indices on the U.S. snow water. Earlier work by Nourani et al. (2018b) evaluated the influence of hydro-climatological variables on the fluctuations of Urmia Lake by applying WTC. Among the vast growing body of literature written on the application of WTC, it has not been used as a method combined with a machine learning skill to downscale the climate models data. Therefore, in this study, WTC, as a data screening method, was employed to select the dominant variables affecting the precipitation and temperature of Tabriz synoptic station. To this end, the coherence between the output of climate models and predictands was investigated. The Coupled Model Intercomparison Project (CMIP) provides climate models which are modern tools to investigate the climate condition in the future (Wilby & Wigley 1997). A host of studies have examined the performance of the CMIP5 models in projecting the climate variables (Kumar et al. 2013; Pendergrass & Hartmann 2014). The CMIP5 offers models under different representative concentration pathways (RCPs; Taylor et al. 2012; Stouffer et al. 2017). The new cutting-edge CMIP6 has recently been provided to compensate for the downside of the CMIP5. CMIP6 models are constantly being updated and comprise higher spatial resolutions than CMIP5 (Eyring et al. 2016). CMIP6 provides a new set of emission scenarios incorporating various shared socioeconomic pathways (SSPs; Riahi et al. 2017). Furthermore, CMIP6 is equipped with more simulations for various scenarios to portray inner inconsistency (Eyring et al. 2016; Stouffer et al. 2017). Numerous research works have investigated the performance of CMIP6 models in simulating the precipitation and temperature in various areas (Akinsanola et al. 2020; Jiang et al. 2020; Rivera & Arnould 2020; Wang et al. 2020; Ngoma et al. 2021).

Ultimately, the objectives of the present study can be listed as follows:

• i.

  To provide a robust wavelet-based predictor screening tool and compare the results with the traditional classic coefficient correlation (CC) method. In this regard, several climate models such as BNU-ESM from CMIP5, and Can-ESM5 and INM-CM5 from CMIP6 were applied.

• ii.

  To feed the screened dominant predictors into the ANN in order to implement downscaling.

• iii.

  To project the future precipitation and temperature based on the ANN model.

Tabriz city (latitude 38.07°N, longitude 46.14°) is the capital city of East Azerbaijan province located in northwestern Iran (see Figure 1). It lies in the valley of the Quru River. Tabriz is surrounded by mountains (e.g., Sahand at 3,707 m) and scattered hills in three directions, namely Pachkin, Eynaly, and Payan. The west side of the city lies within the Tabriz plain, which slopes gently toward the Urmia Lake (60 km to the west). The elevation of the city ranges between 1,350 and 1,600 meters above sea level. The climate of Tabriz is categorized as a continental climate with regular seasons. This gives the city dry and semi-hot summers and cold, wet winters. The city used to receive 360.2 mm of precipitation on average, per year in the past and during the recent decade, it has decreased to 280 mm. The average minimum and maximum temperatures of the city are −2 and 19 °C in January and July, respectively. The average relative humidity varies between 73% in December and 48% in June.

Climate change has led to many predicaments in various parts of the world specifically in arid regions. The city of Tabriz is located in an arid and semi-arid area near the Urmia Lake, the largest salt lake in the region, where the negative impacts of the climate change along with the water resources management policies have dried up the lake. Therefore, an assessment of the future precipitation and temperature variations of the city is required (Zarghami et al. 2011; Baghanam et al. 2020; Sadeqi & Dinpashoh 2020). This can help decision-makers to reach better solutions to climate-related challenges.

In this study, two predictands (i.e., precipitation and temperature) and the outputs of three climate models, as predictors, were applied. Monthly precipitation and temperature of Tabriz synoptic station for the period of 1951–2018, which have been provided by the Tabriz Meteorological Organization, were used in this study. The distribution of monthly precipitation and temperature of Tabriz station are depicted in Supplementary Appendix Figure 1. As shown in the figure, the maximum precipitation fluctuation occurs within April, October, and December indicating the wet season. Accordingly, the driest month takes place during July. In a similar way, the highest and the lowest temperature occur in July and January, respectively.

In order to implement the proposed downscaling method, the predictors from two climate models under the AR6-CMIP6 (Assessment Report 6 and Coupled Model Intercomparison Project datasets) and the predictors under the AR5-CMIP5 (Assessment Report 5 and Coupled Model Intercomparison Project datasets) from the IPCC data distribution center during January 1951 to December 2005 were extracted. Furthermore, future climate variables for CMIP5 models under the RCPs (i.e., RCP4.5 and RCP8.5) and climate variables for CMIP6 models under SSPs (i.e., SSP2 and SSP5) for the future simulation were used in this study. The historical and future data for the period of 1951–2060 were retrieved from https://esgf-node.llnl.gov/. Climate models including BNU-ESM, Can-ESM5, and INM-CM5, which respectively have been developed by research centers in China, Canada, and Russia, were used in this study (Table 1), owing to the preceding research works proving that these three climate models appear to be more effective for investigating the climate of the region (Baghanam et al. 2018; Nourani et al. 2019a, 2019b). Since, in this study, two Earth system models (ESMs) and one GCM were used, henceforth the term climate models (CMs) will be used to refer to both models. Since multiple studies have recommended applying several grid points around the study station will lead to more thoroughgoing results (Frost et al. 2011; Guo et al. 2012; Beecham et al. 2014), in the present study, based on the resolution of CMs, four grid points for each CM were considered nearby the study area (Figure 1). Yet according to the proposed methodology, dominant predictors from potential grid points were selected later on. As depicted in Figure 1, A1, A2, A3, and A4 represent the considered four grid points for BNU-ESM and Can-ESM5, which share the same resolution and their grid points fully overlap. Similarly, B1, B2, B3, and B4 indicate grid points for INM-CM5. Table 1 shows the attributes of CMs and potential variables considered in this study.

The objective of this paper is to carry out a wavelet-based predictor screening method to develop an ANN-based statistical downscaling technique and project the future precipitation and temperature of Tabriz city. Figure 2 demonstrates the blueprint of the study containing three steps. The first step is to determine dominant predictors from the selected CMs by means of feature extraction technique (i.e., WTC and CC). At the second step, dominant predictors screened from the first step were fed into the ANN model to downscale precipitation and temperature. Eventually, at the third step, based on the potential inputs and proper ANN model, which developed during the second step, the projection for the future precipitation and temperature of Tabriz synoptic station under the SSP2 and SSP5 for Can-ESM5 and INM-CM5 and RCP4.5 and RCP8.5 for BNU-ESM was executed. A full description of the proposed methodology is delineated hereunder.

Throughout this step, firstly, dominant periodicities for observed precipitation, temperature, and predictors were separately obtained by CWT. This could bring detailed information that would not have been identified by the original time series. Thereafter, the correlation between predictors and predictands was estimated by WTC and CC. In order to examine the influence of each predictor scale average coherency for each predictor was calculated by extracting the matrix of significant scale (at the 95% confidence level) WTC.

It should be noted that, in the absence of the data pre-processing techniques for selection of the dominant predictors, there would have been 95 large-scale climate variables for BNU-ESM and INM-CM5 and 120 variables for Can-ESM5, meaning that in the aggregate, 1,240 variables should be constituted in the downscaling procedure. If the trial and error method for each parameter had been used, then 2¹²⁴⁰–1 trial and errors would have been required for determining the dominant predictors, which is a time-intensive progress.

WTC and CC used in this step are based on the nonlinear and linear correlation-based methods. This would be a contributing factor in investigating the most influential variables which have a linear or nonlinear relationship with hydrological parameters (i.e., precipitation and temperature). Afterward, selected predictors based on the WTC and CC methods were separately used in the ANN modeling to downscale the climate variables, and the performances of the two methods were evaluated.

In this stage, according to the dominant predictors obtained from three CMs in the first step, an ANN-based downscaling method for statistical downscaling was developed.

CMs often expose substantial deviations from the locally observed data, known as biases. Therefore, to perform downscaling, these biases should be corrected. Standardization is one of the common techniques to eliminate CMs biases, which has been applied to data in several studies (Tripathi et al. 2006; Anandhi et al. 2009). Standardization is implemented by subtracting the mean of observed data and then dividing by the standard deviation.

Ultimately, based on the model developed in the previous step, simulation of the future precipitation and temperature under SSP2 and SSP5 from CMIP6 and RCP4.5 and RCP8.5 from CMIP5 during the period of 2021–2060 was conducted. It should be mentioned that due to the diverse possible perspectives on human-origin effects that influence the environment, such as the rate of population growth, greenhouse gases emission, and economic condition, each RCP and SSP represents a specific radiative forcing pathway. In this respect, SSP2 and RCP4.5 are related to the moderate while SSP5 and RCP8.5 are associated with high emission scenarios.

In order to implement the proposed methodology, the applied mathematical tools were explained briefly in the following sections.

Throughout the recent decades, researchers sought to use a wide range of methods for extracting the features of time series. Among those, Fourier analysis can be mentioned. Nevertheless, Fourier analysis has limitations to be applied to nonstationary time series and can be used for investigating signals which have consistent amplitude and statistical features over time (Kaiser 2011). Thus, Windowed Fourier transform (WFT) was offered to fill this gap (Kaiser 2011). WFT, however, is limited to properly analyzing the time-frequency localization and yielded imprecise and ineffective results. The imprecision stems from its limited capability to identify low and high frequencies, and the ineffectiveness is because of the fact that it has to be done several times according to the scale of the windows to select the best dominant frequencies (Torrence & Compo 1998). Therefore, wavelet analysis as a scale-independent method can be deployed.

In the present study, in order to project the future precipitation and temperature of the Tabriz station within the period of 2021–2060, an ANN-based downscaling method equipped with a data pre-processing wavelet technique was employed. The results of the study are presented in the three steps as follows.

At the first step, WTC was used to determine the dominant periodicities in predictands and predictors, which led to the selection of dominant predictors from three CMs (i.e., BNU-ESM, Can-ESM5, and INM-CM5) based on the coherence method (WTC). To this end, first, the CWT of the two predictands was obtained. Results denoted that there is a 8–16 months periodicity for precipitation throughout 1952–1960, 1936–1968, 1969–1974, 1975–1978, 1980–1982, 1992–1994, and 2001–2004 at 95% confidence level (see Supplementary Appendix Figure 2(a)). Similarly, for average monthly temperature, there is a significant periodicity at the scale of 8–16 months (see Supplementary Appendix Figure 2(b)). Likewise, CWT for the predictors was also employed, which indicated that there is significant periodicity at the periodic scale from 8 to 16 months for the air temperature at grid point 1, geopotential height at grid point 3, eastward wind and geopotential height at grid point 4 (see Supplementary Appendix Figure 3). Subsequently, WTC between the predictors and predictands with meaningful periodic fluctuation was obtained. In order to select the dominant predictors based on the WTC technique, given that the observed precipitation and temperature have significant variability at the periodic scale from 8 to 16 months, a matrix of significance at the 95% confidence level was detached from the WTC values, and SA wavelet coherency of these values for each predictor was ranked. Thereupon maximum values of SA were selected as the dominant predictors. The results of the WTC method for selecting the predictors are tabulated in Table 2.

For the purpose of selecting dominant predictors based on the classic correlation method, CC between the predictands and predictors was calculated. After listing the predictors, the statistical student's t-test (Pu & Ginoux 2017) was applied to identify predictors with meaningful differences. Student's t-test is a helpful method to draw conclusions about the population not just samples. Therefore, predictors with the maximum value of CC were evaluated and selected. According to Table 2, selected dominant predictors for precipitation downscaling, based on the CC technique, were the variables related to humidity, namely relative humidity (hur), specific humidity (hus), and water evaporation flux (evspsbl). The selection of predictors related to humidity variables from grid points A1, B1, A2, B2, and A3 indicates a linear relation with precipitation, which respectively originated from the substantial impact of humidity from the Caspian Sea and Urmia Lake (see Figure 1).

According to Table 2, the representative predictors based on the WTC method yielded different results compared to the CC method. Predictors related to geopotential height (zg), eastward wind (ua), northward near surface wind (vas), and specific humidity (hus) at the determined grid points in Table 2 became dominant according to nonlinear relationship. In this way, the obtained dominant variables from WTC, which have a nonlinear relationship with the precipitation, can represent the selected variables via the linear method (i.e., humidity), owing to the fact that humidity is affected by variability in sea level height, wind velocity, temperature, and air pressure. This is in line with the natural mechanism, meaning that temperature and wind velocity are driving forces to generate humidity. Figure 3 depicts the WTC between precipitation and dominant predictors of BNU-ESM. There are strong anti-phase relationships between predictors and precipitation at 8–16 months periodicity. This shows the anti-phase relationship between the condition of air layers and temperature with the occurrence of precipitation. Figure 4 shows WTC between precipitation and dominant variables of Can-ESM5, where precipitation has a statistically significant positive and negative coherence with the selected predictors according to Table 2. Figure 5 illustrates the WTC between dominant predictors and precipitation for INM-CM5 climate model. Accordingly, based on the selected predictors, strong in-phase and anti-phase coherence are noticeable between the precipitation and the predictors at 8–16 months signal.

For downscaling the temperature, according to Table 2, based on the CC method, the dominant predictors detected for all three CMs at all four grid points were temperature-related large-scale variables (i.e., tas). This indicates a remarkable linear correlation between observed temperature and large-scale CM outputs.

Similar to the precipitation, according to Table 2, WTC for temperature demonstrates different results compared to the CC method. According to the coherence values between predictors and temperature variables, predictors related to height above the sea level (i.e., zg), air pressure (i.e., psl) and temperature (i.e., tas) and near surface humidity (i.e., huss) were identified as dominant predictors. This shows the ability of the WTC to catch nonlinear relationships between variables in addition to linear relationships. Figure 6 illustrates the WTC between temperature and BNU-ESM predictors, where in-phase coherence can be observed between the predictors and temperature. Dominant predictors for Can-ESM5 CM are illustrated in Figure 7. According to the figures, both in-phase and anti-phase relationships can be identified between the temperature and predictors. Figure 8 demonstrates the high in-phase and anti-phase correlation of INM-CM5 predictors with temperature at 8–16 months periodicity.

In the second step, the dominant inputs selected by two feature extraction methods (i.e., WTC and CC) were considered as the input of ANN as a nonlinear downscaling model. In order to put all the data in a comparable range, they were standardized during 1951–2005 in advance. The datasets were divided into two parts: the first 75% of the dominant inputs and observed data which is equal to the period 1951–1991 were used for calibration, and the 25% of them which were equivalent to the period 1992–2005 were applied to validation step. Prior studies have indicated the effectiveness of such data partitioning for the purpose of investigating the hydro-climatological processes (Baghanam et al. 2018; Nourani et al. 2019a, 2019b).

For implementing the ANN downscaling, the three-layer FFNN with back propagation training algorithm and the Levenberg–Marquardt method was used, and tangent sigmoid was considered as the activation function. The optimized number of hidden layer neurons and number of epochs were obtained through the trial and error technique. To this end, the optimum value of evaluation criteria with 360 epochs with 5 neurons in the hidden layer yielded an efficient result. RMSE and DC as the evaluation criteria were calculated for examining the effectiveness of the modeling. The results of this step are tabulated in Table 3. According to Table 3, downscaling results obtained from the WTC predictor screening method outperformed that of the CC technique. Therefore, for precipitation downscaling, WTC method performs 11–16% more effectively based on the DC validation criterion. In the same way, for temperature, downscaling based on the WTC method indicates more satisfactory results, as 35–41% efficiency for DC in validation step was obtained. This improvement appears to be rooted in the ability of WTC to opt for the linear and nonlinear features of the climate data, resulting in a more comprehensive assessment of the impacts of the driving forces on the local climate variables. Hence, the results of the ANN downscaling method based on the WTC are more promising in predicting the future climatic variables. As shown in Figures 9 and 10, it can be inferred that for both precipitation and temperature downscaling, the application of the WTC predictor screening technique outdoes the CC method. Moreover, for all three CMs (i.e., BNU-ESM, Can-ESM5, and INM-CM5), temperature downscaling denotes better performance in terms of evaluation criteria than precipitation. This is in line with the results of previous studies (Sun et al. 2015; Ahmadalipour et al. 2017) which is due to the stochastic nature of the precipitation and deterministic nature of the temperature that tends to have a steady and unvarying quality.

Throughout the final step, the projection of the future precipitation and temperature based on the ANN optimal downscaling model was performed to predict the future oscillations (i.e., 2021–2060) under RCP4.5 and RCP8.5 scenarios for BNU-ESM and SSP2 and SSP5 scenarios for Can-ESM5 and INM-CM5 CMs.

Given the results of the second step, as the downscaling based on the WTC predictor screening method indicated better results than the CC method, the projection of future climate variables (i.e., precipitation and temperature) was performed in accordance with the dominant variables selected by the WTC technique. To this end, large-scale future climate variables for each CM were used to implement the ANN simulation model. For the purpose of evaluating the ANN projection, the variation for projected monthly precipitation and average temperature and observation data at the historical period (i.e., 1951–2005) is illustrated in Figures 11 and 12 for each season of the year. According to Figure 11, there will be both rise and fall in the seasonal future precipitation of the study area. While a decrease of precipitation will be a trend in winter, summer, and fall for all three CMs under both scenarios, spring will experience a climb in precipitation. The underlying cause of this phenomenon can stem from the climate change effects due to the global warming. To elucidate, during the cold months (e.g., January, February, and December), air pollution can be an obstacle for condensation. Pollutant particles can block the incoming sunlight to the earth, thus, the differential temperature on the earth surface and the atmosphere would not be adequate for water vapor to grow and form clouds. On the other hand, in spring, it is observed that due to the decrease of contaminants, which is mostly because of the wind coming from Tabriz city northeastern (Gholampour et al. 2014), the pollution can be scattered. This mechanism will be accompanied by an increase in the air temperature, giving impetus for the formation and growth of clouds. The most decreasing trend in the future precipitation can be seen in Can-ESM5 under the low emission SSP2 scenario, which is 28.92%. Furthermore, BNU-ESM under the high emission RCP8.5 scenario revealed a 7.1% decrease, which is the minimum value.

As is shown in Figure 12, a growing trend for temperature is noticeable for three CMs, while RCP4.5 and SSP2 show mild increase, the growth of temperature under the RCP8.5 and SSP5 is remarkable. Furthermore, according to the difference between the predicted and historical values of temperature, it can be concluded that INM-CM5 under the high emission SSP5 scenario shows 4.21 °C growth in temperature, which is the highest increase for the region in the future. The minimum rate of increase is attributed to BNU-ESM under the low emission RCP4.5 scenario, numbering 2.18 °C. It should be taken into consideration that, despite the decreasing and increasing trends respectively at precipitation and temperature values, the climate of the study region will experience more extreme weather events. This is closely aligned with the Intergovernmental Panel on Climate Change 2014 report, which remarked that ‘the rate of heavy precipitation and the recurrence of the heatwave is expected to go up in most areas, and Asia in particular’.

This paper studied the impacts of climate change on the variation of hydrological variables (i.e., precipitation and temperature) of the Tabriz synoptic station for the period of 2021–2060. In this regard, monthly precipitation and the average temperature of the study area during 1951–2005 were used as the observation period. Three CMs, including BNU-ESM from CMIP5 models, and Can-ESM5 and INM-CM5 from the CMIP6 models were employed to implement ANN-based statistical downscaling. There are several large scale variables that can influence the climate of the area; however, the extent of this impact is not equal for all the variables, while some predictors show a strong relationship with the predictands, others do not show a significant connection. Therefore, in order to increase the accuracy of the ANN downscaling model, it is crucial to select the dominant inputs to avoid data redundancy. In this regard, WTC and CC as data-dimensionality reduction methods were used for selecting the potential variables, which led to predictor screening. The result of the ANN downscaling models indicated that the WTC predictor screening approach outperformed the CC method, due to the capability of WTC to distinguish the nonlinear relationship between predictors and predictands. WTC indicates 11–19% improvement for modeling the precipitation and 41–45% for temperature in the base period (1951–2005) in comparison to the CC approach. Finally, the future projection of precipitation and temperature was implemented and results denoted a decreasing trend in the precipitation, while the temperature will experience an increasing rate. This condition is testament to the fact that the galloping rate of GHGs emissions will lead to more global warming, and it consequently has adverse effects on the precipitation rate. All in all, it is expected that the precipitation will undergo a 7.1–28.92% decrease, while the temperature will experience a 2.18–4.21 °C increase.

In this study, a wavelet-based data screening method equipped with an ANN model was used to implement statistical downscaling of CM outputs. It is recommended to use other predictor screening methods such as supervised approaches (e.g., PCA) and unsupervised approaches (e.g., clustering) to examine the performance of different methods. Furthermore, it is suggested to incorporate more AI-based models (e.g., SVM, ANFIS, and GEP) or other statistical downscaling approaches (e.g., MLR, SDSM). This can contribute to more recognition of the capabilities and weak points of different models in downscaling the climate variables. Moreover, using different bias correction methods such as Delta Change and Quantile Mapping methods could be investigated in further research works. It is proposed to utilize the ensemble of different AI models to test and compare their accuracy with the single models. Using several CMs and more scenarios could also give a more comprehensive insight into the future climate variation in a region.

All relevant data are available from an online repository or repositories https://esgf-node.llnl.gov/projects/esgf-llnl/.