https://orcid.org/0000-0003-3647-3507
Abstract
The purpose of this study is the projection of climate change's impact on the Groundwater Level (GWL) fluctuations in the Mashhad aquifer during the future period (2022–2064). In the first step, the climatic variables using ACCESS-CM2 model under the Shared Socio-economic Pathways (SSPs) 5–8.5 scenario were extracted. In the second step, different machine learning algorithms, including Multilayer Perceptron Neural Network (MLP), Adaptive Neuro-fuzzy Inference System Neutral Network (ANFIS), Radial Basis Function Neural Network (RBF), and Support Vector Machine (SVM) were employed for the GWL fluctuations time series prediction under climate change in the future. Our results point out that temperatures and evaporation will increase in the autumn season, and precipitation will decrease by 26%. The amount of evaporation will increase in the winter due to an increase in temperature and a decrease in precipitation. The results showed that the RBFNN model had an excellent performance in predicting GWL compared to other models due to the highest value of R² (R² = 0.99) and the lowest value of RMSE, which were 0.05 and 0.06 meters in training and testing steps, respectively. Based on the result of the RBFNN model, the GWL will decrease by 6.60 meters under the SSP5-8.5 scenario.
HIGHLIGHTS
The CMhyd model was used to extract climatic variables from the ACCESS-CM2 model.
Temperatures and evaporation will increase and rainfall will decrease.
The Radial Basis Function Neural Network model had an excellent performance in GWL prediction.
The GWL will decrease by 6.60 m under the SSP5–8.5 scenario in the Mashhad aquifer.
INTRODUCTION
Due to the increased population and the growth of urban and rural areas, amenities, and agricultural requirements, the amount of demand for water consumption has increased. It is provided through surface and groundwater resources. Reduced rainfall during the hot seasons (Dai et al. 2020) and poor surface water quality are some of the reasons for water supply using natural resources. Groundwater resources are used as natural sources for various purposes like drinking and irrigation in different regions (Çadraku 2021).
Due to the emission of greenhouse gases such as CO2 caused by human activities, including the use of fossil fuels, the earth is warming (Montzka et al. 2011). Long-term changes in temperature and weather patterns that are caused by natural or unnatural factors are called climate change (United Nations). According to the Intergovernmental Panel on Climate Change (IPCC) report, the world temperature will increase by 1.5°C by 2050, which will cause poverty and pose dangers in the lives of more than 100 million people. Avoiding it requires a significant reduction in carbon dioxide emissions before 2030 (Masson-Delmotte et al. 2018). Based on the United Nations reports, climate change, directly and indirectly, causes extensive changes in different regions. Its most critical direct effects are the global increase in temperature (Lionello & Scarascia 2018; Arnell et al. 2019; Boukal et al. 2019) and extreme fluctuations in rainfall that cause drought in some regions (Berg & Sheffield 2018; Cook et al. 2018; Dai et al. 2018; Goodarzi et al. 2019; Kim & Jehanzaib 2020; Philip et al. 2020). It causes severe floods in some regions (Hirabayashi et al. 2013; Arnell & Gosling 2016; Hettiarachchi et al. 2018; Swain et al. 2020). Among the indirect effects of climate change, we can point out the economic consequences (Ciscar et al. 2011; Park et al. 2018; Allam & Jones 2019; De Angelis et al. 2019; Piontek et al. 2021), social (Schwindt et al. 2016; Markkanen & Anger-Kraavi 2019; Austin et al. 2020), changes in marine ecosystems (Blanchard et al. 2012) and eco-geomorphology of wetlands and beaches (Day et al. 2008), creating extensive changes in the physical characteristics of the soil (Trnka et al. 2013), changes in genetic plants (Alsos et al. 2012), and changes in the forest ecosystem (Price et al. 2013) and plant species (Thuiller et al. 2011). Also, it changes the GWL by changing the aquifer's recharge (Guevara-Ochoa et al. 2020).
As mentioned earlier, temperature changes and weather patterns play a significant role in the intensity and amount of rainfall. Considering that groundwater resources are formed from surface and subsurface water infiltration into the depths of the earth, changes in the intensity and amount of rainfall have a direct role in them. Since groundwater resources have a vital role in the water supply and increasing global warming, a projection of the climate change impacts on the GWL fluctuations is necessary. Investigating it is done through software and numerical modeling. Some of the studies in this field are mentioned below.
The HadCM3 model was used to investigate climate change impacts on the Umm er Radhuma unconfined aquifer in the western plains of Iraq from 2020 to 2099 (Hassan 2020). The effect of climate change on surface and groundwater resources was investigated using the combination of the Soil and Water Assessment Tool (SWAT) and Modular Finite-Difference Flow (MODFLOW) models and the Coupled Model Intercomparison Project – Phase 5 (CMIP5) – in the Yom and Nan basins in northern Thailand (Petpongpan et al. 2020). Their results showed that under the Representative Concentration Pathway (RCP) 2.6 and 8.5 scenarios, the average annual temperature will increase by 0.5–0.6 and 1–0.9 °C, respectively. The effect of climate change on the groundwater resources was studied in the Lake Tana basin in Ethiopia (Tigabu et al. 2021). Their results indicated that the flow rate of groundwater resources would decrease compared to the base period. The research results regarding the effect of climate change, urbanization development, and sea level change on groundwater resources under the A1B, B1, and A2 scenarios showed that the GWL will decrease (Akbarpour & Niksokhan 2018). The effect of climate change on the GWL was studied under RCP2.6, RCP4.5, and RCP8.5 scenarios in the Shabestar Plain in Iran (Jeihouni et al. 2019). The study results regarding the effect of climate change on groundwater recharge using the CORDEX model under the RCP scenarios showed that the amount of precipitation will decrease by 20% and the evaporation will increase under the RCP4.5 scenario by 8.1% in the Tekeze basin (Kahsay et al. 2018). Also, the amount of groundwater supply will decrease. Khoi et al. (2022) investigated the effects of climate change using the CMIP6 model and the SWAT model in Ho Chi Minh City in Vietnam.
Machine learning algorithms have capabilities in hydrology and water resource management (Yaseen et al. 2019; Tao et al. 2022). Easy implementation, low cost, and high accuracy of artificial intelligence (AI) algorithms can be mentioned (Morshed-Bozorgdel et al. 2022). In this regard, Anaraki et al. (2021) studied the effect of climate change on flood frequency using hybrid machine learning methods. Their results indicated that machine learning methods had acceptable performance results in the downscaling of precipitation and temperature. Kadkhodazadeh et al. (2022) studied reference Evapotranspiration (ETo) prediction under climate change based on machine learning, multi-criteria decision-making, and Monte Carlo methods. Their results showed that the TOPSIS method was the best algorithm, while the accuracy of machine learning algorithms was very good for ETo modeling. El Bilali et al. (2021) used machine learning algorithms, including Adaptive Boosting (Adaboost), Random Forest (RF), Artificial Neural Networks (ANNs), and Support Vector Regression (SVR) algorithms, to predict groundwater quality for irrigation purposes. Based on the reviews published by Rajaee et al. (2019) and Ahmadi et al. (2022), AI methods perform well in simulating and predicting GWL fluctuations in different aquifers.
In the lack of hydrogeological information conditions, the ANN is a helpful tool to predict the GWL (Lee et al. 2019). However, El Bilali et al. (2021) results showed that the performance of Adaboost and RF models was higher than that of ANN for groundwater quality prediction. Various types of research have been conducted to predict the GWL fluctuations using the ANN, which can be referred to as the research of Poursaeid et al. (2022), Samani et al. (2022), Seidu et al. (2022), and Singh & Panda (2022). Kenda et al.’s (2018) results indicated that the Gradient Boosting algorithm was the best performance in comparison with another algorithm for GWL prediction in the Ljubljana Polje aquifer. Kenda et al. (2018) used different machine learning algorithms, including Linear Regression (LR), Decision Trees, RF, and Gradient Boosting models for the GWL prediction in the Ljubljana Polje aquifer.
The effects of climate change were investigated under RCP2.6, RCP4.5, and RCP8.5 scenarios in the Tasuj plain in Iran using the ANN, Least Square Support Vector Machine (LS-SVM), and the Nonlinear Autoregressive Exogenous Model (NARX) (Ghazi et al. 2021). Farzin et al. (2022) investigated the Ground Resource Index (GRI) and Groundwater Table (GWT) prediction using a hybrid of the Bi-directional Long Short-Term Model (BLSTM) and the Harris Hawk Optimization (HHO) algorithm, namely the BLSTM HHO algorithm, the Long Short-Term Model (LSTM), ANN, Seasonal Autoregressive Integrated Moving Average (SARIMA), and the Autoregressive Integrated Moving Average (ARIMA) algorithms. The effect of climate change on the groundwater resources in the Gaza Strip was investigated (Al-Najjar et al. 2021). Jeihouni et al. (2021) investigated the effects of climate change on the GWL in the Shabestar Plain under the RCP2.6, RCP4.5, and RCP8.5 scenarios during 2020–2050. The deep learning method was used to project GWL in Germany until 2100 under climate change (Wunsch et al. 2022). The results of using the ANN and NARX models indicated a decrease in GWL. Seifi et al. (2020) used ANN, SVM, and Adaptive Neuro-Fuzzy Inference System (ANFIS) algorithms for GWL modeling in the Ardabil plain in Iran. Farzin & Valikhan Anaraki (2021) investigated suspended sediment load prediction under climate change patterns using the combination of the hybrid LS-SVM and Flower Pollination Algorithm (FPA) in comparison to ANFIS methods. The results of GWL projection using LR, SVM, Gaussian Process Regression (GPR), and Neural Network (NN) models showed that the performances of GPR and LR models are the best (Sapitang et al. 2021). Sahoo et al. (2017) investigated GWL changes in agricultural regions of the USA based on machine learning algorithms.
Due to population growth in recent years and climate change in arid and semi-arid regions, the lack of rainfall, and the reduction of surface water flows, the projection of the climate change impact on the GWL is vital in the management and control of water resources in the future. Although according to the conducted research, climate change has a significant effect on the quality and quantity of groundwater resources, no research has been done to predict climate change under the IPCC Sixth Assessment Report (AR6) on the GWL fluctuation time series in the Mashhad aquifer. Since the Mashhad aquifer is one of the critical basins in Iran due to the passage of important rivers like the Kashafroud River and it has faced severe water resource crises, such as floods and droughts in recent years, the projection of GWL fluctuations under climate changes in future periods plays an essential role in creating the attitude of water resource management in this region. The clear purposes of this research are as follows: (1) using the ACCESS-CM2 model from the AR6 under the SSP5–8.5 scenario to predict climatic variable changes, including maximum and minimum temperatures, precipitation, and evaporation during the future period (2022–2064) compared to the historical period (1992–2021) and (2) projection of the GWL fluctuations in the Mashhad aquifer using different machine learning algorithms, including Multilayer Perceptron (MLP), ANFIS, Radial Basis Function Neural Network (RBFNN), and SVM. The results of our research will play a significant role in the control and management of groundwater resources in this region.
METHODS AND MATERIALS
Study area
Location of the Mashhad aquifer.
Groundwater resources
The Mashhad aquifer is an unconfined aquifer with an area of 3,351 km2. The thickness of its layer fluctuates by 130 m. The morphology of the bedrock in this area is very uneven and has appeared on the alluvial surface. In some places, the thickness of alluvium reaches up to 300 m. The characteristics of the Mashhad watershed are shown in Table 1.
Characteristics of Mashhad watershed
| Characteristics | Value |
|---|---|
Watershed area ( ) | 9,909.4 |
| Precipitation average (mm) | 247.5 |
| ETo average (mm) | 2,300 |
| Elevation average (m) | 1,214.3 |
| Minimum elevation (m) | 900 |
| Maximum elevation (m) | 1,600 |
| Characteristics | Value |
|---|---|
| Watershed area () | 9,909.4 |
| Precipitation average (mm) | 247.5 |
| ETo average (mm) | 2,300 |
| Elevation average (m) | 1,214.3 |
| Minimum elevation (m) | 900 |
| Maximum elevation (m) | 1,600 |
According to the Khorasan Razavi Regional Water organization (www.khrw.ir), the number of groundwater sources in the Mashhad plain is 7,433, including springs, aqueducts, deep wells, and semi-deep wells. There are 6,000 wells in the Mashhad plain, which the volume water extracted is more than 4 billion cubic meters. The annual discharge of water resources is 1,116.27 million cubic meters, which 850.30 million cubic meters are used for agriculture, 32.11 million cubic meters for industry, and 233.86 million cubic meters for drinking purposes.
Water balance
In the water resource balance cycle, the relationship between different balance factors and input and output sources has been established. The inputs of the water resource balance cycle include rainfall, surface and groundwater flows, and transfers. The outputs of this cycle include evaporation, exploitation of surface and groundwater resources, evaporation from lakes, surface, underground outflows, and transfers. The annual rainfall volume in the Mashhad plain is 2,704.97 million cubic meters, and the annual evaporation volume from rainfall is 2,246.77 million cubic meters. Therefore, the effective rainfall is 458.2 million cubic meters. The amount of incoming and transfer surface flows is 99.36 million cubic meters. Therefore, the total volume of water produced in the Mashhad plain is 557.56 million cubic meters. The total volume of water surface and groundwater utilization in the Mashhad plain is 1,124.78 million cubic meters. The volume of its net consumption is 607.32 million cubic meters. The volume of water surface and groundwater flow in the Mashhad plain is 28.24 and 11.68 million cubic meters, respectively. The volume of evaporation from the lakes is 11.3 million cubic meters in this plain. Therefore, the volume of water resources extracted in the Mashhad plain is 658.54 million cubic meters. Considering the volume of water consumption and water supply mentioned in the Mashhad plain, the volume of water balance is 100.98 million cubic meters in the Mashhad plain.
Groundwater level fluctuations
Monthly time series: (a) temperature; (b) evaporation; (c) rainfall; and (d) GWL during the historical period in the Mashhad aquifer.
Monthly time series: (a) temperature; (b) evaporation; (c) rainfall; and (d) GWL during the historical period in the Mashhad aquifer.
Methodology
Flow chart of the study.
Datasets
Characteristics of the General Circulation Model (GCM) model used
| Model name | Institution | Country | Grid size (lon × lat) | Reference |
|---|---|---|---|---|
| ACCESS-CM2 | Commonwealth Scientific and Industrial Research Organization and Bureau of Meteorology | Australia | 144×192 | Bi et al. (2020) |
| Model name | Institution | Country | Grid size (lon × lat) | Reference |
|---|---|---|---|---|
| ACCESS-CM2 | Commonwealth Scientific and Industrial Research Organization and Bureau of Meteorology | Australia | 144×192 | Bi et al. (2020) |
Downscaling
In this research, Climate Model data for hydrologic modeling (CMhyd) was used for downscaling climatic data from the General Circulation Model (GCM) model and correcting their bias. The CMhyd tool is written in Python 2.7 using NetCDF41, NumPy, and SciPy packages and in the PyQt42 program environment. Bias correction methods include linear scaling, delta change correction, precipitation local intensity scaling, power transformation of precipitation, variance scaling of temperature, and distribution mapping of precipitation and temperature. According to the research results by Ringard et al. (2017), Switanek et al. (2017), Smitha et al. (2018) and Enayati et al. (2021), the distribution mapping method has a good performance for bias correction. This technique was used to correct the accuracy of the simulations resulting from the microscaling of GCM models and to minimize the difference between the observed and simulated climate variables (Rathjens et al. 2016; Vaittinada Ayar et al. 2021). In this study, after selecting the investigated climate variable in the CMhyd model environment and selecting the type of bias correction method, introducing the input path of the investigated climate variable, which were downloaded in netCDF (*.nc file) format from the ACCESS-CM2, introducing daily observational climate data (1992–2021) in ASCII format, climate data were extracted in the form of text file for the future period.
Machine learning
is the normalized variable value, X is the main value, 
and 
are the maximum and minimum values, and 
and 
are the maximum and minimum ranges for normalizing the investigated variable. Two series of data should be created to perform the calibration and validation steps to ensure the efficiency of each machine learning method. The purpose of machine learning is to find the most appropriate weight vector, bias vector, and minimize the error function. After the training step, the validation stage is performed. At this stage, the model is executed from another part of the data that was not used in the network training stage. In this step, the model outputs are compared with the observed values. Finally, the accuracy of the model is calculated. The training and validation steps are repeated until finding the best result from the model. Finally, the test stage is performed for the final control of the model. Machine learning algorithms used in this research include ANFIS, SVM, MLP, and RBFNN. MATLAB software was used for this purpose. The general structure of machine learning NN methods used in the present study is shown in Figure 4. In the following, a summary of the performance basis of different machine learning algorithms will be mentioned. 
Machine learning methods’ architecture of the present study.
Multilayer Perceptron Neutral Network
is the weight in the hidden layer, 
is the input variable, 
is the bias for the hidden neuron j, and 
is the output variable:
Radial Basis Function Neural Network
is the weight of the edges, b is the bias, and f is the Gaussian activation function.
Support Vector Machine
In the above equation, 
and 
are the Lagrange coefficients and 
is the kernel function.
Adaptive Neuro-fuzzy Inference System Neural Network
An ANFIS that approximates real continuous functions has good performance in training, generation, and classification. This model was introduced by Jang (1993). Using this model, fuzzy rules can be extracted from numerical data. The structure of the ANFIS model consists of five layers, including input nodes, rule nodes, intermediate nodes, subsequent nodes, and output nodes.
Evaluation of Machine Learning Models
) and root mean square error (RMSE), were used. The 
and RMSE values close to 1 and 0 indicate the high accuracy of the model to predict the studied variable. The evaluation criteria equations used in this research are shown in the following:
In the above equations, 
, 
, and 
show the observed, predicted, and the average of the predicted values, respectively.
RESULTS AND DISCUSSION
Results of downscaling
Figure 5 shows the annual average changes of rainfall variables and minimum and maximum temperatures during 2022–2064 under the SSP5-8.5 scenario. Based on it, the precipitation area will have the highest amount in the northwestern and western regions of Iran. The southeastern regions in Iran will have the lowest amount of precipitation with an annual rainfall of 63 mm/year. The amount of precipitation in the investigated area, located in the northeastern regions, is the average amount in Iran.
Yearly changes in precipitation, minimum, and maximum temperatures under the SSP5–8.5 scenario during 2022–2064.
Yearly changes in precipitation, minimum, and maximum temperatures under the SSP5–8.5 scenario during 2022–2064.
Monthly percentage changes in climatic variables. (a) Minimum and maximum temperatures and (b) rainfall and evaporation under the SSP5–8.5 scenario during 2022–2064 compared to the historical period.
Monthly percentage changes in climatic variables. (a) Minimum and maximum temperatures and (b) rainfall and evaporation under the SSP5–8.5 scenario during 2022–2064 compared to the historical period.
The performance of different machine learning algorithms
Monthly time series for (a) minimum and maximum temperatures and (b) rainfall and evaporation during 2022–2064.
Monthly time series for (a) minimum and maximum temperatures and (b) rainfall and evaporation during 2022–2064.
Changes in the parameters of GWL predictive models.
According to Figure 9, RBFNN, ANFIS, MLP, and SVM models have the best performance in the test stage.
The evaluation of the agreement between the actual and predicted values using (a) RBF, (b) MLP, (c) ANFIS, and (d) SVM algorithms.
The evaluation of the agreement between the actual and predicted values using (a) RBF, (b) MLP, (c) ANFIS, and (d) SVM algorithms.
Monthly GWL prediction during 2022–2064 using different machine learning algorithms.
Monthly GWL prediction during 2022–2064 using different machine learning algorithms.
Monthly GWL prediction compared to the historical period.
The most significant decrease will occur in the summer and autumn seasons. It is the consequence of the decrease in precipitation and the increase in temperature and evaporation in the summer and autumn seasons under the SSP5–8.5 scenario based on Figure 11. Therefore, the decrease in the GWL in the future period can be related to the increase in temperature and evaporation. Ghazi et al. (2021) found that the GWL would decrease in the Tasuj plain in Iran. Our result is consistent with the results of Chang et al. (2015), Arkoç (2022), and Wunsch et al. (2022).
CONCLUSIONS
The purpose of this research was to determine the effect of climate change on the GWL fluctuations in the Mashhad aquifer from 2022 to 2064. In this research, the historical climate from the synoptic station and GWL datasets from the WRM of Mashhad were used. Also, the ACCESS-CM2 model under the SSP 5–8.5 scenario was used to generate climatic variables in the future. The CMhyd model was used to downscale the climatic data from the ACCESS-CM2 model under the SSP 5–8.5 scenario. Different machine learning algorithms, including MLP, ANFIS, RBFNN, and SVM, were employed to predict the GWL fluctuations under climate change in the future periods.
According to the results of predicted climate variables in the Mashhad aquifer using the CMIP6 under the SSP5–8.5 scenario, the highest increasing minimum and maximum temperature changes in the future period (2022–2064) will belong to October by 2.50 and 3.16 °C, respectively. Moreover, the amount of precipitation will decrease by 26% in the future. The amount of evaporation will increase in the autumn season. The findings of this research indicated that the most amounts of precipitation will be unavailable due to the increase in temperature and evaporation in the future, which will hurt the feeding rate in the Mashhad aquifer. The results of various machine learning algorithms, including RBFNN, MLP, ANFIS, and SVM models to predict the effects of climate changes on the GWL of the Mashhad aquifer, showed that the RBFNN method has the best performance in projecting the GWL fluctuations compared to the other algorithms due to the highest value of R2 (R2 = 0.99) and the lowest value of RMSE, which were 0.05 and 0.06 m in the training and testing steps, respectively. The result of GWL predicted under climate change showed that it will decrease by 6.60 m in this region. Since the changes in weather patterns in the long term lead to the reduction of the GWL, it causes an increase in desert areas. Reducing these resources for agricultural, industrial, and domestic purposes will face a severe crisis. Therefore, the improved management of the Mashhad aquifer is vital to control the water demand and supply. The findings of this research have a practical role in making helpful groundwater resource management decisions.
We have obtained satisfactory results; however, some limitations have been observed. Some of the fluctuations observed in the GWL changes are related to the aquifer dynamics in the Mashhad aquifer. Since we did not consider the aquifer dynamics in machine learning algorithms due to the lack of information, models with a hydrogeological basis, like MODFLOW, should be used. The accuracy of machine learning models should be evaluated by considering the aquifer dynamics, the complexities, and the types of uncertainties using MODFLOW in future work. Considering the different behaviors of GWL fluctuations in dry and wet seasons, the accuracy of machine learning algorithms will be investigated in different time scales in future work. Moreover, it is suggested that further investigation be conducted regarding using underground dams to reduce evaporation and the relationship between the drought index and the salinity level of the Mashhad aquifer in future periods.
AUTHOR CONTRIBUTIONS
All authors contributed to the study's conception and design. M.H.E. contributed to the supervision and implementation of the research, and to the analysis of the results of the manuscript. G.P. designed the figures. The first draft of the manuscript was written by M.H.E. and G.P. Material preparation, data collection, and analysis were performed by G.P., M.H.E., and A.R. Editing of the manuscript was done by A.F. and S.R.K. All authors read and approved the final manuscript.
DATA AVAILABILITY STATEMENT
All relevant data are included in the paper or its Supplementary Information.
CONFLICT OF INTEREST
The authors declare there is no conflict.
%20cropped.png?versionId=7143)