Improving the statistical downscaling performance of climatic parameters with convolutional neural networks

SDSM

Statistical downscaling model

ANN

Artificial neural network

CNN

Convolutional neural network

AI

Artificial intelligence

GCM

General circulation model

MI

Mutual information

RF

Random forest

ML

Machine learning

CC

Correlation coefficient

RMSE

Root mean squared error

DC

Coefficient of determination

RELU

Rectified linear unit

SSP

Shared socioeconomic pathways

QM

Quantile mapping

EQM

Empirical quantile mapping

Conv

Convolutional

Assessing future climate variability has become a significant challenge in developing mitigation and adaptation strategies in the environmental field. Although general circulation models (GCMs) are widely employed to forecast future climate conditions, their coarse resolution data often lack the detail required for accurate impact assessment and informed decision-making regarding climate change. Reconstructing historical climate is crucial for assessing climate variability and informing policy decisions related to climate change. To overcome this limitation, the utilization of reanalysis data and downscaling techniques can offer valuable advantages. This enables the accurate reconstruction of historical climate and assessment of climate variability at finer scales. The downscaling techniques offer several significant advantages, including the ability to enhance spatial resolution, which is crucial for capturing localized climate features such as temperature gradients, precipitation patterns, and wind behavior. By incorporating local topographic and land surface characteristics, downscaling methods can improve the accuracy of climate projections, enabling more reliable assessments of climate impacts on various sectors such as agriculture, water resources, and infrastructure. In addition, downscaling facilitates informed decision-making by providing policymakers, planners, and stakeholders with detailed and actionable climate information for adaptation and mitigation strategies. Numerical climate data often contain systematic biases, which can reduce the accuracy of the modeling output. To address this issue, various bias correction methods are available, ranging from simple techniques such as linear scaling to more sophisticated methods like quantile mapping (QM) (Shrestha et al. 2017), where each technique is suited for individual predictors such as precipitation, temperature, and solar radiation (Gudmundsson et al. 2012a). Thus, for precipitation and temperature downscaling, linear scaling, daily translation, local intensity scaling, power transformation, distribution mapping, and QM methods are being used as bias correction methods. Many such bias correction methods use a predefined cumulative distribution function (CDF) to create a correction function (Piani et al. 2010). Danandeh Mehr & Kahya (2017a, 2017b) employed a novel bias correction method to mitigate systematic biases by scaling the long-term mean of observed and simulated annual rainfall maxima. Moreover, this procedure improved the accuracy in reproducing frequency-based statistics of observed rainfall series at each grid point. Furthermore, various bias correction techniques exhibit both advantages and drawbacks. For instance, Chen et al. (2013) highlighted that linear scaling fails to address alterations in the precipitation frequency distribution. Conversely, Danandeh Mehr & Kahya (2017a, 2017b) showcased that the grid-based, pattern-oriented approach offers the benefit of reduced computational expenses by utilizing monthly datasets to align long-term averages of observed and simulated data. This is in contrast to traditional methods, which typically rely on daily observations to correct GCM/regional climate model (RCM) outputs on a daily basis. Moreover, Chen et al. (2013) demonstrated that bias correction methods based on distribution consistently outperform those based solely on the mean.

There are two primary categories of downscaling: statistical and dynamical downscaling. Dynamical downscaling involves using RCMs, which are driven by boundary conditions of climate fields such as pressure, wind, temperature, and humidity from GCMs (Zhang et al. 2020). However, this approach requires significant computational power, expertise, and a large volume of input data (Baghanam et al. 2019). The expertise required entails knowledge in model setup, atmospheric physics, and boundary layer processes. In addition, proficiency in data analysis, model validation, and uncertainty quantification is crucial for assessing model performance and interpreting downscaled projections effectively. For example, in addressing uncertainties arising from both model and parameters, Dai et al. (2024) utilized two distinct approaches. The first approach, based on variance analysis using Sobol indices, primarily focuses on model output and within-model variance. Meanwhile, the second approach, employing a moment-based method with multi-model Averaged Moment Analysis (AMA) indices, integrates between-model variance effects through a model choice term. In this way, to compensate dynamical downscaling drawbacks, statistical downscaling methods may be preferred. Statistical downscaling involves using statistical-based computations to establish a relationship between predictors and local-scale predictands (i.e., in situ data) and to bridge the spatial scale gap between climate parameters (Benestad et al. 2008). This approach relies on historical predictors and local variables to establish statistical relationships without requiring knowledge about the underlying physical processes (Mosavi et al. 2018). In addition, a hybrid statistical–dynamical downscaling model that combines the benefits of both statistical and dynamical downscaling could be also applied (Boé et al. 2006; Sun et al. 2015; Ma et al. 2022). This approach can capture the fine-scale dynamics while also benefiting from the computational savings of a statistical model (Walton et al. 2015; Wang et al. 2019). Statistical downscaling encompasses a wide range of techniques, from simple regression models to complex artificial intelligence (AI) methods. There are three main subcategories of statistical downscaling approaches: (i) Linear regression methods, such as the statistical downscaling model (SDSM); (ii) Weather generators, such as the Long Ashton Research Station Weather Generator (LARS-WG) (Racsko et al. 1991) and (iii) Nonlinear regression methods, such as AI-based methods. According to the literature, linear-based downscaling methods may not perform well when dealing with nonlinear climatic relationships (Dorji et al. 2017). Nonlinear models have demonstrated higher accuracy compared to other statistical downscaling techniques (Campozano et al. 2016). AI-based downscaling techniques include a variety of machine learning (ML) methods, such as support vector machines (Chen et al. 2010; Selvanathan et al. 2023), random forests (Ooi et al. 2022), relevance vector machines (Ghosh & Mujumdar 2008), artificial neural networks (ANNs) (Nourani et al. 2018), and gene expression programming (Hashmi et al. 2011). The new generation of ANNs, called deep learning (DL) techniques, which include convolutional neural networks (CNNs), generative adversarial networks, and long short-term memories (LSTMs), have recently been applied in climate modeling. Although DL models are more complex and difficult to interpret than shallow models, they have shown to outperform shallow models in downscaling tasks (Wang et al. 2021a, 2021b; Nourani et al. 2023).

One major challenge with statistical downscaling is the quality of the input data. To overcome this issue, preprocessing methods such as predictor screening can be used to improve the accuracy of the model output by eliminating redundant information and reducing the convergence time. Alternatively, if the input data are insufficient, data augmentation is employed to enhance the data in quantitative terms (Wei et al. 2022). Various supervised and unsupervised methods such as random forests (RF) (Pour et al. 2016), least absolute shrinkage and selection operator, and mutual information (MI) have been used to extract dominant predictors. Filter-based methods, such as the correlation coefficient (CC), can identify linear relationships between predictors and predictands. However, they are not able to detect complex relationships. Embedded methods, such as RF, can result in the most accurate output by explicitly optimizing the attribute set. Wrapper methods, such as backward stepwise regression, can outperform the correlation coefficient method and rely on information about feature relevance obtained from classification tasks (Hessami et al. 2008). Moreover, hybrid methods combine filters and wrappers to leverage the strengths of each approach. Shongwe et al. (2006) utilized statistical downscaling and screening techniques, including principal component regression and canonical correlation analysis. Their findings indicated that both predictor screening methods exhibited comparable performance due to the inherent structural similarities. The efficacy of these methods in reducing multicollinearity and improving screening efficiency was demonstrated. As a DL method, CNNs have demonstrated the ability to effectively extract features from high-dimensional input data, making them a promising technique for downscaling applications (Baño-Medina et al. 2020a). CNNs have the potential to address the time-consuming and multicollinearity problems associated with traditional shallow plain models, allowing for more efficient and accurate downscaling results (Sun & Lan 2021).

Tabriz, situated amid Iran's diverse landscapes, faces multifaceted challenges stemming from the effects of climate change. Over recent years, the region has witnessed a notable uptick in temperatures, resulting in more frequent and intense heatwaves. In addition, shifts in precipitation patterns have brought about irregularities in rainfall distribution, impacting the availability of water resources critical for agriculture, industry, and daily life. These climatic shifts have heightened the risk of both drought and flooding events, further straining the city's infrastructure and exacerbating concerns over water scarcity. Moreover, the changing climate poses significant threats to public health, with rising temperatures increasing the prevalence of heat-related illnesses and exacerbating respiratory conditions. This study employed both deep and shallow learning methods to downscale precipitation and temperature of Tabriz city. The term ‘shallow learning’ refers to ML models that utilize a single hidden layer, such as feedforward neural networks (FFNNs) in this study (Wang et al. 2021a, 2021b). The reason for utilizing shallow plain AI-based models relies on assessing the performance of novel DL models, coupled with various preprocessing techniques. Shallow learning is performed using FFNN standalone and coupled with predictor screening methods, namely, RF, MI, and CC. DL is performed using CNNs. Historical input data are drawn from the ERA5 reanalysis at four grid points around the Tabriz synoptic station, Iran. The biases are corrected using empirical quantile mapping (EQM), and the results of downscaling were compared with the widely used linear SDSM. Ultimately, future precipitation and temperature were projected under two shared socioeconomic pathways (SSPs), i.e., SSP2-4.5 and SSP5-8. The most significant novelty of this paper lies in the integration of both EQM and RF methods into the preprocessing stage for CNN-based downscaling. Given that the study's scope is focused on projecting future climate change, temperature and precipitation are recognized as two primary drivers of climate change (Wu et al. 2022). Therefore, the study aimed to focus on these two parameters.

The study aimed to accomplish the following objectives:

• (i)

  Evaluate the newly proposed RF predictor screening for the selection of dominant predictors.

• (ii)

  Comparison of linear and nonlinear predictor screening methods.

• (iii)

  Comparison of classic and DL-based ANNs for downscaling of climatic parameters.

• (iv)

  Comparison of future climate projections.

• (v)

  Assessment of the impact of climate change scenarios under SSP2-4.5 and SSP5-8.5.

The precipitation pattern in Tabriz exhibits seasonal variations, with the highest levels occurring during the spring months, followed by a decrease during the summer. Tabriz is influenced by both cold northern winds from Siberia and moist western winds from the Black Sea. These winds play a crucial role in shaping the city's climate, resulting in cold, snowy winters and relatively mild springs. Situated in a mountainous region, Tabriz experiences variations in temperature and precipitation, attributable to the interplay of these winds and its geographical location.

For this study, daily time series data were collected from the Coupled Model Intercomparison project Phase6 (CMIP6) GCM data spanning the base period from January 1979 to December 2010. The observation data for precipitation and temperature were obtained from the local meteorological organization and used as the predictand data for the analysis. Furthermore, the dataset was split into training, validation, and test subsets based on the period of the data. The data from 1979 to 2000 were used as the training set, data from 2000 to 2006 were used as the validation set, and the rest as the test set. Furthermore, to obtain an accurate downscaling model, the daily ERA5 reanalysis datasets, which have a spatial resolution of 0.75° × 0.75°, were mapped by linear interpolation onto the dimensions of the GCM data to be applied for future scenarios. To predict future trends, two future datasets were utilized: SSP2-4.5 and SSP5-8.5. These datasets represent different emission scenarios: while SSP2-4.5 indicates a more realistic scenario, SSP5-8.5 represents a high emission scenario. The grid points around the study station are distributed in different regions in which the first and second grid points are situated in Iran, but the third and fourth grid points are situated in Armenia and Azerbaijan, respectively (see Figure 1). The collected GCM dataset contains various atmospheric variables, including humidity, pressure, wind, temperature, divergence, airflow, zonal and meridional flow, and vorticity, as shown in Table 1, which are used as input parameters in the models referred to as ‘predictors’. Their purpose is to establish a relationship between these parameters and observations.

In this study, the EQM technique was employed to address systematic biases in precipitation predictors. EQM utilizes the empirical CDF derived from local and ERA5 time series data. By correcting biases in the ERA5 dataset based on the mean and variability differences between ERA5 and observed values over the reference period, EQM enhances the accuracy of precipitation predictions. To overcome overfitting and reduce the dimensionality of data, predictor screening methods of the RF, MI, and CC were utilized. The RF and MI methods helped select relevant predictors by identifying information and constructing optimal decision trees.

The decision tree method due to its interpretability and ability to capture nonlinear relationships has gained popularity in time series issues. RF is an algorithm used for classification and regression proposed by Breiman (2001). RF is an ensemble model that builds a forest of trees with randomly selected features, which performs with high accuracy on high-dimensional data. By selecting trees, this method produces unbiased error estimates and improves classification performance by selecting features, randomly.

MI or information gain is a measure rooted in entropy theory, which quantifies the information between two random variables using Shannon entropy calculations (Shannon 1948). In technical terms, a rise in MI signifies an escalation in the importance of the associated variables.

In this stage, imposing the dominant predictors, which were extracted in the first step, then used as inputs to the downscaling models, which were developed using both FFNN as a shallow learning and CNN as a DL method. Their applied structures are briefly explained as follows.

The convolutional layer is located immediately after the input layer, and its output is fed to the fully connected layer. The convolutional layer is responsible for extracting important features as feature maps through the convolution of the input dataset with a learned kernel, while assuming a nonlinear activation function.

To project future mean temperature and precipitation values over the study area, two emission scenarios, SSP2-4.5 and SSP5-8.5, were utilized as reference scenarios. SSP2-4.5 represents a realistic emission scenario, considering the current development situation without additional adaptations or extra emissions. In addition, SSP5-8.5 was used as a projecting future climate scenario, representing high-rate fossil fuel development society and high-rate population increase, reflecting the worst-developed society in terms of emissions. To incorporate the emission scenarios, dominant predictors selected from step 1 were considered inputs for future projection. The projection procedure involved employing the optimum network as a model, with SSP variables serving as inputs. Since there are no existing predictands in this step, the weights and biases of the optimum network were used as a reference to project future climate changes.

The primary objective of this study was to predict future mean temperature and precipitation patterns over Tabriz city in Iran for 2010–2100. Specifically, nine distinct downscaling models were employed in this study including FFNN, FFNN-RF, FFNN-MI, FFNN-CC, CNN, CNN-RF, CNN-MI, CNN-CC, and SDSM. Future projections were implemented based on two SSPs, including SSP2-4.5 and SSP5-8.5.

The obtained results according to the proposed methodology are presented in the following subsections.

As shown in Figure 9, the Taylor diagram was generated for four grid points located around the study area to compare the performance of the bias-corrected model. The position of the plotted point on the diagram represents the performance of the model at a particular grid point. Based on the Taylor diagram, it was observed that the second grid point exhibits a better match with Tabriz station compared to the other three grid points. The plotted point for the second grid point indicated a higher correlation between the bias-corrected data and observations, and had a smaller standard deviation, indicating better agreement between the bias-corrected data and observations.

Following bias correction, predictor screening was carried out using the MI, RF, and CC methods. Table 2 presents the results of this analysis, which identified the most highly ranked predictors as prominent features. These predictors were subsequently selected as inputs for downscaling models. Initially, CC analysis was applied to quantify the linear relationship between predictor variables and the target variables of precipitation and temperature. Subsequently, MI analysis was employed to assess the nonlinear dependencies between predictors and target variables, capturing intricate associations beyond linear relationships. Finally, RF methodology was implemented to comprehensively evaluate the importance of predictors by considering their collective influence on the prediction outcomes.

The analysis revealed that the most influential predictors based on the screening methods were those related to temperature and geopotential height. The relationship between temperature and precipitation in Tabriz is multifaceted and can be influenced by various factors, including regional climate patterns, topography, and atmospheric circulation. In Tabriz, the surrounding temperature (i.e., temperature at all grid points) significantly impacts precipitation, primarily through the process of evaporation from nearby bodies of water, such as the Caspian Sea, Sevan, and Urmia Lake. As temperatures rise, the rate of evaporation increases, leading to higher atmospheric moisture content and favorable conditions for cloud formation. During winter, cold air from surrounding mountains flows into Tabriz, resulting in increased snowfall and precipitation due to the lower moisture-holding capacity of colder air. In addition, in warmer months, high temperatures in adjacent areas trigger the formation of low-pressure systems, drawing moisture from these bodies of water and contributing to the development of thunderstorms and other forms of precipitation.

Geopotential height is an important predictor of local precipitation that is nonlinearly connected to local precipitation. Geopotential height is a measure of the actual height of a pressure surface above mean sea level. Therefore, when measuring the geopotential height in a specific location, it provides an indication of the height of the pressure surface at that location. On the other hand, low-pressure systems are characterized by a lower central pressure than their surrounding regions. When the winds meet the low pressure, the air rises in the atmosphere. In the study area, which is surrounded by water bodies such as Urmia Lake, Sevan Lake, and the Caspian Sea, rising air can lead to the condensation of water vapor and the formation of clouds, which often results in precipitation. Furthermore, the study area is affected by different southwesterly troughs such as the Red Sea and Eastern Mediterranean troughs. These troughs transfer a significant amount of humidity to the study area, which can lead to instability. This instability can further influence local weather patterns and precipitation levels.

Based on the RF predictor screening approach, four meteorological variables were chosen as significant predictors for precipitation: geopotential height, wind direction, divergence, and specific humidity. Among these, geopotential height emerged as the most dominant predictor.

Divergence is a fundamental concept in meteorology that represents the net outflow of air. As such, it has been found to be closely associated with local precipitation patterns, with the degree of correlation being nonlinear in nature (Watson & Blanchard 1984). Specifically, divergence patterns are typically linked to upper-level troughs and ridges, which are regions of the atmosphere where the air is either rising or sinking. In particular, upper-level divergence often occurs in conjunction with upward motion in the troposphere's interior, which can lead to cloud formation and precipitation at downstream of the upper-level ridge. The Caspian Sea region, where the divergence predictor is located, is known for its high levels of atmospheric moisture. As the air in this region is lifted by the upper-level divergence, it is expected to cool and expand, leading to the formation of clouds and precipitation. Moreover, stronger divergence forces air to rise more vigorously, which can intensify the precipitation patterns in the region and the air transfers those patterns to the study region.

The influence of wind on precipitation patterns is a crucial aspect of meteorological studies. Extensively researched, wind serves as the primary agent for transporting atmospheric parameters across different regions. Consequently, it plays a pivotal role in shaping the distribution of moisture, temperature, and other meteorological variables. In particular, the wind predictor at the fourth and second grid points is expected to carry moisture content from the Caspian Sea toward the study area, which can have a significant impact on local precipitation patterns. As such, the wind that blows across the sea is often laden with water vapor, which can be carried inland and contribute to the formation of clouds and precipitation.

In the context of the CC method, the primary predictors influencing the precipitation factor are identified as mean sea level pressure and divergence. Furthermore, airflow strength and precipitation are observed to emerge as dominant predictors, and their relationship is inherent in the method, displaying linear correlation with temperature.

This study investigated five downscaling approaches for climate data; a shallow FFNN standalone, FFNNs coupled with a predictor screening method, and DL method without any predictor screening approach. The efficiency of these newly proposed models was evaluated using the SDSM as a benchmark method.

All of the inputs of the downscaling models were split into 70, 20, and 10% parts for training (from 1979 to 2000), verification (from 2000 to 2006), and test (from 2006 to 2010), respectively. AI-based models were developed using the Levenberg–Marquardt optimization algorithm and the rectified linear unit (RELU) activation function. The training dataset obtained from the screening step and bias-corrected ERA5 dataset were used as inputs in the first layer to the FFNN and CNN models, respectively. The second layer, which corresponds to hidden layers, stabilized the relationship between inputs and target data, and the last layer led to the outputs. The optimum number of hidden neurons and epochs were obtained as shown in Table 3, where the number of epochs and hidden neurons guarantee the speed of training, convergence, and accuracy of the model, obtained through the trial-and-error procedure. The number of inputs and neurons in Table 3 indicates the dominant predictors extracted by RF, MI, CC, and without the screening procedure. By adjusting the hidden layers of shallow networks by just one layer, there is a noticeable impact on performance due to their limited depth, whereas adding extra layers yields marginal improvements. Conversely, in deep networks, adjusting the number of layers significantly influences performance, with additional layers often leading to enhanced results as they enable the network to capture increasingly complex patterns and hierarchies of features.

According to the obtained results presented in Tables 4 and 5, the CNN-QM model exhibited superior performance when considered both as a standalone model and when coupled with various predictor screening methods. In addition, it was concluded that the models that did not inherit from the QM-based bias correction method performed poorly. Consequently, it was determined that the CNN method obviates the need for a separate predictor screening step, primarily due to the presence of the convolutional layer, which automatically extracts the most dominant predictors. As outlined in Tables 4 and 5, nonlinear predictor screening techniques outperformed their linear counterparts. This superior performance of nonlinear predictor screening methods can be attributed to the inherent stochastic nature of the predictand. On the other hand, the classical FFNN method coupled with predictor screening methods showed inferior performance.

According to Figure 10(a), it can be noted that the CNN-QM effectively captures temperature fluctuations with negligible biases in this modeling approach. In addition, Figure 10(b) illustrates the coupling of CNN-QM with RF predictor screening, resulting in weaker outcomes compared to standalone CNN-QM. The rationale behind this diminished performance lies in the CNN architecture's requirement for more extensive input data, which RF diminishes by reducing the input data size. Figure 10(c) depicts the CNN-QM coupled with the MI predictor screening procedure. The accuracy of this modeling approach is inferior to RF-based CNN, indicating that MI performs weaker than RF in screening climatic variables as predictors. Ultimately, Figure 10(d) illustrates CNN-QM with linear predictor screening, resulting in the weakest performance due to the linear assumption applied to input predictors.

The accuracy of the conducted models is visually represented in Figure 11, where the downscaled temperature predictand is illustrated through a series of scatter plots. Specifically, Figure 11(a) presents the best-fitted scatter plot, followed by Figure 11(b)–11(d), each showcasing subsequent analyses.

Similar to temperature results, for precipitation, the time series often exhibit random fluctuations, which can be challenging for modeling techniques. In Figure 12(a), it is evident that the CNN-QM method effectively captures these fluctuations, showcasing its ability to handle the inherent variability in precipitation data. However, Figure 12(b) illustrates a slight decline in performance when CNN-QM is integrated with RF predictor screening. This decline could be attributed to the inherent limitations of RF in capturing the patterns present in precipitation data, potentially leading to an incomplete representation of the underlying dynamics. Moving to Figure 12(c), where CNN-QM is combined with the MI predictor screening procedure, a similar pattern of diminished accuracy compared to the RF-based CNN approach is observed. This outcome suggests that while MI attempts to address the complexity of precipitation data, it may not adequately discern the relevant predictors, resulting in suboptimal model performance. Finally, in Figure 12(d), where CNN-QM is coupled with linear predictor screening, the weakest performance is evident. This could be attributed to the oversimplification imposed by linear screening methods, which may fail to capture the nonlinear relationships inherent in precipitation data.

Compared to shallow-layered networks, CNN downscaling offers several advantages. Specifically, CNNs can effectively train the downscaling process of predictands in deeper layers, automatically extracting features without requiring manual screening. In addition, leveraging batch normalization and larger batch sizes result in faster training times. These findings suggest that DL-based downscaling using CNNs has the potential to improve the accuracy and efficiency of downscaling procedures.

However, it is expected to see an increase in precipitation during the spring season. The underlying cause of this phenomenon is dependent on various factors. One of the potential factors is climate change, which is expected to alter the precipitation patterns, resulting in drier conditions in already dry regions and wetter conditions in already wet regions. As Tabriz city is a semi-arid region, it is likely to experience a drier climate. Moreover, the fluctuations in precipitation trends during spring and other months can be attributed to air pollution. Air pollution has the ability to disrupt the overall moisture balance in the atmosphere, leading to reduced precipitation during the winter season. Conversely, during spring, the decrease in contaminants allows for their dispersion. This process is accompanied by a rise in atmospheric temperature, which will provide impetus for the creation and expansion of clouds, which finally leads precipitation.

As depicted in Figure 14(b), all seasons are projected to display an increasing trend by the end of this century. The findings of the future temperature projection conducted under SSP2 indicate an anticipated temperature rise of 1.31 °C. Notably, the most substantial temperature increase is projected under SSP5 by an amount of 2.9 °C. The principal factor contributing to these divergent outcomes lies in the development of the SSP scenarios. SSP5 assumes the continuation of existing trends, characterized by a significant reliance on fossil fuels for energy generation. Consequently, the elevated temperature projection under SSP5 can predominantly be attributed to the heightened levels of greenhouse gas emissions resulting from increased utilization of fossil fuels. Furthermore, SSP5 incorporates assumptions of higher population growth, which drives urbanization, amplifies industrial activity, and consequently leads to increased greenhouse gas emissions, ultimately exacerbating global warming.

In this study, the efficiency of shallow and DL-based downscaling methods, including classical FFNN and CNN, were evaluated for projecting future precipitation and temperature using ERA5 data from four surrounding grid points around the Tabriz synoptic station. To remove systematic biases, the EQM method was utilized as a bias correction technique. Furthermore, the second grid point, situated in the vicinity of the large water body known as the Caspian Sea, emerged as the most influential grid point for precipitation formation. Predictor screening methods of RF, MI, and CC were then employed to identify the dominant predictors and find the relationship between predictors and predictands. In the predictor screening approach, the RF method assumes nonlinear relationships between ERA5 climatic data and predictands. It selected pressure, humidity, and temperature-type predictors, demonstrating superior performance compared to other predictor screening methods. On the other hand, the CC method, which employs mean sea level pressure as the predictor for precipitation and utilizes airflow strength and zonal velocity as predictors for temperature, while assuming linear relationships, demonstrated the least effective performance when compared to the other predictor screening methods. This helped remove extra features and overcome overfitting of the FFNN model. Subsequently, FFNN and CNN downscaling models were developed and compared using RMSE and DC for the temperature and precipitation predictands in both training and validation sets. The results showed that the CNN-based downscaling model performed better than the FFNN model without predictor screening, yielding lower RMSE and higher DC by the amount of 0.01 mm and 0.71 for precipitation downscaling and 0.08 °C and 0.98 for temperature downscaling, respectively. Finally, two scenarios, SSP2-4.5 and SSP5-8.5, were utilized to project the far future (i.e., 2100) temperature and precipitation variability using the appropriate downscaling model. The results indicated that the most significant increase in temperature and decrease in precipitation occurred under SSP5-8.5 by the amount of 2.9 °C and 3.5 mm, respectively. The obtained results of the study suggested that Tabriz is at risk of encountering prolonged periods of drought, particularly during the dry seasons. This potential scenario is further compounded by a noticeable trend toward warmer winters in the region. These changes indicate a significant departure from historical climate norms and may have far-reaching implications for the local ecosystem, agriculture, and water resources. In addition, there is an alarming signal regarding the possible absence of snowfall in the area. Historically, snowfall has been a vital source of water and has played a crucial role in maintaining the regional water balance. If this trend continues, it could have dire consequences for the region's water supply. To address the imminent challenges posed by climate change in Tabriz, policymakers need to implement a comprehensive strategy. This strategy should include the development of climate-resilient infrastructure, water conservation measures, diversification of water sources, and support for sustainable agriculture. It is recommended to use other nonlinear predictor screening tools, such as supervised principal component analysis and AI-based feature selection methods (i.e., genetic programming) over numerical weather prediction data, to enhance the performance of shallow-layered ANNs. Since this study is based on a single reanalysis dataset, it is recommended that future studies utilize RCMs to address the limitations of the large-scale ERA5 data. In addition, remote sensing data can be valuable for conducting local-scale impact assessments. Furthermore, the use of anomaly detection methods is suggested to identify unusual patterns in the climatic data. Moreover, employing hybrid CNN-long short-term memory (LSTM) downscaling models is recommended for extracting influential predictors using CNN and enhancing prediction capabilities using LSTM. Finally, other environmental factors can be incorporated into the modeling procedure as drivers of climate change. These factors include solar radiation, soil moisture, and wind patterns.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.