Potential impacts of climate change on groundwater levels in Golpayegan Plain, Iran

In arid and semi-arid regions, population growth and spatio-temporal limits of the availability of surface water have led to overextraction of groundwater resources, resulting in a sharp decline in water levels in aquifers. In recent years, climate change has added to the stress on these sources by affecting groundwater recharge sources (decreasing precipitation and runoff and rising temperatures) and depletion of aquifers (Bell et al. 2014; Eslamian 2014; Zamanirad et al. 2018). Located in an arid region, Iran is also facing limited water resources and water deficits. Meanwhile, periodic droughts and climate change have led to the overextraction of aquifers and a sharp decline in groundwater levels (Mohammadi Ghaleni & Ebrahimi 2015).

An efficient strategy for managing groundwater resources is to assess the effects of different parameters on groundwater level. Changes in groundwater levels by effective factors can be predicted through modeling, which can be done using different conceptual, physical, numerical, and statistical methods (Rajaee et al. 2019).

Numerous studies have employed Artificial Intelligence models for groundwater-level prediction considering their advantages such as simplicity of use, high accuracy, and applicability to aquifers based on limited data. These efforts include using Artificial Neural Networks (Coulibaly et al. 2001; Daliakopoulos et al. 2005; Nourani et al. 2008; Rakhshandehroo et al. 2012; Kouziokas et al. 2018), Neuro-Fuzzy Systems (Jalalkamali et al. 2011; Kholghi & Hosseini 2009; Shirmohammadi et al. 2013; Khaki et al. 2015), Genetic Programming (Fallah-Mehdipour et al. 2013; Shiri et al. 2013; Sadat-Noori et al. 2020), Support Vector Machine (Yoon et al. 2016; Nie et al. 2017; Mukherjee & Ramachandran 2018), and hybrid models. Emadi et al. (2021) and Ahmadi et al. (2022) performed modeling of groundwater for Iran and all over the world, respectively.

Some of the main factors involved in simulating and predicting groundwater levels are the selection of the model structure, number, type, time scales, and model inputs (Moravej et al. 2020). Even the choice of input data will affect the accuracy of prediction results (Sadat-Noori et al. 2020). Spatio-temporal analysis of groundwater levels can also significantly affect the modeling accuracy (Tang et al. 2019).

Some studies show that the recharge and discharge of groundwater depends on precipitation, climatic variables, human impacts such as agricultural activities and construction of weirs for flow regulation (Allen et al. 2004; Woldeamlak et al. 2007; Taylor et al. 2013). Hence, predicting the recharge and discharge for future climatic change conditions is of great importance for integrated water management (Patil et al. 2020). Extensive research has been conducted in recent years to assess the effects of climate change on groundwater resources. Most of these studies are focused on the effects of climate change on groundwater recharge (Scibek et al. 2007; Goderniaux et al. 2009; Lee et al. 2019; Patil et al. 2020; Petpongpan et al. 2020; Azizi et al. 2021; Ouyang et al. 2021) or changes in groundwater levels (Jeihouni et al. 2019; Patil et al. 2020; Shrestha et al. 2020).

Land subsidence is an environmental hazards for which various mechanisms are effective in its formation and development. From 1925 onwards, many cases of land subsidence around the world have been studied and evaluated (Jackson et al. 2004). Golpayegan plain in Iran, especially the eastern part of this plain, is one of the areas that face this hazard, due to which residential houses and farms located in this area are constantly broken and cracked. Janat & Ghazezadeh (2007) performed a study on land subsidence due to groundwater extraction and presented a report associated with this phenomenon which clarified the relationship between extraction the groundwater and land subsidence in this plain.

Therefore, this study aims to investigate how much climate change can effect monthly groundwater level in the Golpayegan aquifer under the SSP5-8.5 and SSP1-2.6 climate scenarios. For this purpose, this study adopted the CanESM5 model from the IPCC's Sixth Assessment Report for the near future (2017–2032). The Least Squares Support Vector Regression (LSSVR) model was employed with a structure compatible with input data including precipitation, temperature, pan evaporation, soil moisture, and groundwater levels.

It is worth noting that what sets this study apart from prior works includes:

• 1.

  In most of the previous studies, climatic variables have been used as input data to the models. In this research, in addition to climatic data, a soil moisture variable has been used as input to the models

• 2.

  Data-mining methods such as multivariate regression and clustering have been used in order to achieve more accurate results in the present study.

• 3.

  No research has been done in the past to evaluate the effects of climate change in the Golpayegan aquifer.

• 4.

  The assessment of the effects of climate change in previous research is mostly based on the Fifth Report, while in the present study, the scenarios of the Sixth Report have been used to evaluate the effects of climate change on the groundwater level of the Golpayegan aquifer.

With an area of 3,508 square kilometres, Golpayegan is located in the western part of Isfahan province and the southern part of the Namak Lake catchment area in Iran. The area of the unconfined and homogeneous Golpayegan aquifer is approximately 461 square kilometres with an annual discharge volume of about 245 million cubic metres. This aquifer has different hydrogeological characteristics due to its variable morphological and geological conditions. Accordingly, the alluvial thickness varies between 20 and 180 metres. Also, its transfer coefficient is estimated as between 100 and 4,000 square metres per day (Iranian Ministry of Energy 2018).

The location of meteorological and hydrometric stations and observation wells in Golpayegan as well as an aquifer groundwater-level graph for the last 25 years (1993–2018) is presented in Figure 1. Over the 25-year period (1993–2018), it has had a 25 m (1 m annual average) drop in groundwater levels. Groundwater-level changes in the Golpayegan aquifer indicate different fluctuations in the three periods of 1993–2001, 2002–2007 and 2008–2018. Groundwater level has changed in these three periods by −11.9, +0.6 and −18 metres, respectively (Iranian Ministry of Energy 2018). The foremost factors contributing to the declining groundwater levels in the Golpayegan aquifer are the overdrafting (discharge) of groundwater resources and reduced groundwater recharge following the construction of Golpayegan and Baghkol-Khansar dams. After the construction of Kucheri reservoir dam (2008) and the overdrafting in some observation wells (19 and 22), the water-level drawdown in the Golpayegan aquifer reached about 48 m over 16 years (2002–2018).

As outlined in the research flowchart (Figure 2), the research steps involved pre-modeling stages, modeling of the historical period (using the LSSVR algorithm), and predicting the future period. First, the data on precipitation, temperature, pan evaporation, and the water level in the observation wells were collected and the future climatic and soil moisture data were downloaded and downscaled. Then, based on the similarities in the behavior of the observation wells, the aquifer was divided into four clusters and one well was selected as a representative for each cluster. Finally, after modeling and validation in the historical period, the LSSVR model was used to predict the groundwater level in the future period.

The research datasets included the following:

• 1.

  Monthly temperature and precipitation from meteorological stations (Golpayegan, Vaneshan, Robatmaleki, Laybid, Ghalehbabmohammad, Kochary, Hasanabad, Sarabhandeh, and Abbasabad) and monthly groundwater level from 27 observation wells were provided by the Iranian Meteorological Organization and Regional Water Company.

• 2.

  Daily soil moisture was obtained from ESA CCI (https://www.esa-soilmoisture-cci.org/) for the 1978–2018 period at a spatial resolution of 0.25° and a daily temporal resolution (Dorigo et al. 2015). Then, monthly soil moisture was calculated.

• 3.

  The monthly average precipitation and temperature were extracted from the IPCC's Sixth Assessment Report, Global Climate Models (GCMs) of Coupled Model Intercomparison Project 6 (CMIP6) for the period of 2017–2032. This study used the SSP1-2.6 and SSP5-8.5 scenarios.

Missing values and outliers in the mentioned variables were replaced and corrected using statistical methods. In order to prepare the data for modeling, they were normalized and randomly split into 70% for training and 30% for testing. Finally, based on the available time-series data and the possibility of creating a 12-month lag time between input and output data, the study considered a 15-year period (2002–2017) for the precipitation, temperature, soil moisture, pan evaporation, and groundwater-level input variables and a 16-year period (2002–2018) for the output variable of groundwater level.

As shown in Figure 1, there are 27 observation wells in this area, which are classified based on spatial analysis of groundwater-level changes. This analysis aims to identify the similar behaviors of observation wells, cluster them, and select a representative observation well from each cluster. Using the AGglomerative NESting (AGNES) hierarchical clustering method (Nielsen 2016), observation wells 4, 8, 19, and 20 were selected as representatives for clusters 1, 2, 3, and 4, respectively.

In the present study, the groundwater level in the historical period was modeled using the Least Squares Support Vector Regression (LSSVR) model in the MATLAB software. Suykens & Vandewalle (1999) proposed LSSVR (for regression), LSSVM (for classification), and LSSVC (for clustering). LSSVR converts the nonlinear relationship between inputs and outputs to a linear relationship. The advantages of LSSVR include high precision and accuracy, low complexity, mathematical tractability, and speed. One of the factors affecting LSSVR accuracy is the selection of an appropriate kernel function. In this study, the Radial Basis Function (RBF) was investigated.

First, the model was trained and then applied to the validation data and the water-level values were predicted. Finally, these values were compared with the observed data and the results were analyzed.

Based on the suitable model results in the historical period, this model was used to predict the groundwater level in the future period. Input data of the future period were prepared based on the following steps.

In this study, three models from the IPCC's Sixth Assessment Report, BCC-CSM2-MR, CanEsm5, and MIROC6, were used to generate future climatic data as presented in Table 1. After assessing and comparing the values of these models with the observation data, the CanEsm5 model was selected for the final production of future data based on its minimum error and maximum correlation for precipitation and temperature parameters under the SSP1-2.6 and SSP5-8.5 scenarios (Table 2).

Given that only temperature and precipitation data are available under the climatic model and scenarios for the future period, other data such as pan evaporation and soil moisture should be predicted for this period. For this purpose, first, using multiple linear regression in TableCurve 3D, the pan evaporation and soil moisture values in the historical period were considered as functions of precipitation and temperature, and fitted equations were obtained. TableCurve 3D can provide equations for a set of fitted procedures along with statistical parameters such as standard error and coefficient of determination and the values of constant coefficients used in the equations. Then, the values of pan evaporation and soil moisture were predicted in the future period based on these equations and the predicted values of precipitation and temperature by climatic scenarios as independent variables.

Due to the 12-month delay in the groundwater level compared with the inputs and the temporal continuity of the historical and future periods, the level of the first 12 months was also obtained from the average monthly balance of the last five years of the historical period. Using the SSP1-2.6 and SSP5-8.5 climatic scenarios in each of the four clusters resulted in a total of eight scenarios for the future period.

In Equations (3)–(5), and are the measured and predicted groundwater levels, respectively, and and are the average measured and predicted groundwater levels in different observation wells. Based on Equation (5), the negative and positive values of the MBE criterion indicate underestimation and overestimation in the model results, respectively. The best values obtained for the NRMSE, MBE, and R² criteria are 0, 0, and 1, respectively, indicating that the simulated and measured values are equal.

The results are divided into three sections, namely pre-modeling, historical, and future period modeling. These categories are further explained below.

The results of data preprocessing before modeling are presented in this section. Data processing includes spatial analysis (clustering), statistical analysis of parameters, and correlation analysis between the output and input data. The observation wells of the Golpayegan aquifer were clustered to select the observation wells for each cluster based on the AGNES method, as presented in Figure 3.

After clustering, the observation wells in the Golpayegan aquifer were divided into clusters 1 through 4 (Figure 3). There were seven wells (1, 2, 3, 6, 8, 9, and 11) in Cluster 1, four wells (4, 5, 7, and 10) in Cluster 2, eight wells (12, 14, 15, 16, 17, 19, 21, and 22) in Cluster 3, and eight wells (13, 18, 20, 23, 24, 25, 26, and 27) in Cluster 4. Based on criteria such as (1) similarity in water level changes in the representative well compared with other observation wells in each cluster, (2) suitable spatial distribution at the aquifer, and (3) selection of different groundwater behavior patterns, four observation wells, namely 8, 4, 19, and 20, were selected as observation wells in clusters 1 through 4, respectively. The statistical characteristics of the input variables including precipitation, mean temperature, pan evaporation, and soil moisture for each cluster are presented in Figure 4.

According to Figure 4(a), the maximum and minimum precipitation values related to clusters 4 and 2 with 13 and 10 mm per month, respectively. Precipitation in the 90th percentile of the months from Oct 02 to Oct 17 in clusters 1–4 was, respectively, less than 67, 49, 54, and 55 mm. The maximum range of variations in average monthly temperature was −7.4 °C to 30.3 °C in Cluster 4 (Figure 4(b)), whereas the highest median in the average temperatures was seen in Cluster 1 (14.8 °C).

As shown in Figure 4(c), the maximum amount of evaporation was 465.5 mm (Cluster 2) and the maximum average evaporation was 161.4 mm (Cluster 4). According to Figure 4(d), the range of variations in soil moisture was from a minimum of 0.08 (Cluster 4) to a maximum of 0.33 cubic metres of water per cubic metre of soil (Cluster 3). During the study period (2002–2017), soil moisture in clusters 2 and 3 with a median of approximately 0.20 was higher than in clusters 1 and 4 with a median of 0.16 cubic metres of water per cubic metre of soil.

Considering that the random coefficients of the model change after every iteration, the modeling results were extracted from the average of ten consecutive runs and presented as the final model result. The similarity between the range of variations in the simulated and observation variables indicates the accuracy of the model in simulating the output variable variations. The observed and simulated groundwater level changes for the four clusters are presented in Figure 5.

The observed groundwater levels in clusters 1–4 were, respectively, 1,768.8, 1,780.5, 1,726.9, and 1,759 m, and the simulated median values in these clusters were, respectively, 1,769.8, 1,782.1, 1,726.9, and 1,759.1 m. The results show the acceptable accuracy of the LSSVR model in simulating the range of variations in groundwater level in different clusters. The training, testing, and overall values of R², NRMSE, and MBE in the LSSVR model for the four clusters under study are presented in Table 3.

According to Table 3, the highest groundwater simulation accuracy with NRMSE was equal to 0.02 (excellent accuracy) in Cluster 4. The simulation accuracy in Cluster 3 was good (0.10 < NRMSE < 0.20) and moderate (0.20 < NRMSE < 0.30) in clusters 1 and 2. The maximum and minimum coefficients of determination between the observed and simulated water levels in clusters 3 and 2 were, respectively, 0.99 and 0.89 for the entire data. The positive MBE values in all clusters indicate the LSSVR model's overestimation of the groundwater level simulation. This overestimation is, however, negligible and less than 0.06 m for all clusters in the entire dataset. The values and distribution of errors for different clusters are presented in Figure 6.

Figure 6(a)–6(d) shows the nearly normal distribution of errors in clusters 1–4. Moreover, the limited range of error variations in Figure 6(d) (−1.0 < error < 1.0) indicates the highest accuracy of the LSSVR model in simulating the groundwater level in Cluster 4.

The monthly temperature in the near future (2017–2032) compared with the historical conditions has increased under both scenarios, with the highest increase seen in May (0.3 °C) and the lowest in November (0.1 °C) in the SSP5-8.5 scenario. However, precipitation does not have a clear pattern under both scenarios. In the pessimistic scenario (SSP5-8.5), May had the highest increase (23%) and January the highest decrease (56%) in precipitation.

Using TableCurve 3D, in the historical period, pan evaporation (E) and soil moisture (M) values were fitted as a function of precipitation (R) and temperature (T). Their estimation equations were obtained as shown in Table 4. These relationships were used for predicting pan evaporation and soil moisture in the future period from the values of precipitation and temperature predicted by the climatic scenarios.

Based on adjusted coefficient of detection (Adj R² < R²), which considers the degree of freedom (Akbari et al. 2018), suitable correlation was observed. Low fit standard error (FitStdErr) showed low error. Regression relationships were very significant based on the Fstat indicator (Mohseni Movahed et al. 2018).

The variations in groundwater levels in the historical period (2002–2017) and the near future (2017–2032) for different clusters under the SSP1-2.6 and SSP5-8.5 scenarios are presented in Figure 7.

As shown in Figure 7, the groundwater level in the SSP5-8.5 scenario experiences a greater decrease compared with SSP1-2.6 in all clusters. The maximum and minimum differences between the groundwater level at the end of the future period under the SSP1-2.6 and SSP5-8.5 scenarios are related to clusters 3 (Figure 7(c)) and 4 (Figure 7(d)) and are equal to 3.7 and 9.9 m, respectively. This is reasonable in light of the 44.8 m and 4.5 m decline in observed groundwater levels during the historical period in clusters 3 and 4, respectively.

Groundwater is the world's largest freshwater resource and due to overextraction, levels have declined in many regions causing extensive social and environmental impacts. Hence, this study set out to predict groundwater level (GWL) under climate change at a basin scale. It should be noted that effective groundwater management strategies require a deep characterization of GWL fluctuations, in order to identify individual behaviors. Clustering has frequently been applied in hydrology and geohydrology studies to distinguish similar groups of behavior such that each group can be related to different responses of a complex aquifer under any external change. Various researchers like Bloomfield et al. (2015) and Naranjo-Fernández et al. (2020) have used clustering to classify GWL hydrographs. To achieve this purpose, first the spatial analysis of groundwater levels was investigated using the AGNES clustering method, which is one of the clustering techniques. The finding revealed four different behaviors in the observation wells (27 wells) of the Golpayegan aquifer.

Then, future climatic data were extracted from CMIP6 under the SSP1-2.6 and SSP5-8.5 climate scenarios. As a result, unlike precipitation, the temperature has an increasing trend of about 0.2 °C and 0.3 °C in the near future compared with the historical period under the SSP1-2.6 and SSP5-8.5 scenarios, respectively, as in Mishra et al. (2020).

Given the mentioned finding, historical and future period modeling were performed separately in each of the four observation wells from each cluster.

The modeling results for the historical period showed that the highest accuracy in groundwater-level simulation was obtained for Well 20 in.Cluster 4 based on minimum NRMSE (0.02), highest R² (0.93), and lowest MBE (0.024). Groundwater-level changes at the beginning and end of the historical period (2002–2017) in clusters 1–4 are equal to −20.6, −20.3, −44.8 and −4.5 metres, respectively.

Modeling of the future period under climate change scenarios showed a decreasing trend in the groundwater level in all clusters, which was greater under SSP5-8.5 than under the SSP1-2.6 scenario. The values of groundwater-level changes in the two scenarios SSP1-2.6 and SSP5-8.5 at the beginning and end of the near-future period (2017–2032) in clusters 1–4 are equal to 1.4, 5.7, −8.9, −1.9 and −3.4, −0.1, −18.8, −5.6 metres, respectively. In other words, the average monthly drop in groundwater level in the SSP1-2.6 and SSP5-8.5 scenarios is 0.5 and 3.9 cm per month, respectively. Several studies such as Zamanirad et al. (2018) and Jeihouni et al. (2019) confirm the decline in groundwater levels due to climate change in various parts of Iran, which is consistent with the results of this study. In addition to the simulation model, projected precipitation and temperature variations of this study area corroborate that runoff and recharge are decreasing and as a consequence the groundwater level is also declining under climate change.

Therefore, the future groundwater conditions, especially in countries such as Iran, which have an arid and semi-arid climate and are affected by climate change, are especially vulnerable, and so adaptive solutions must be developed for the optimal management of water resources.

Conflict of Interest None.

Available on request.

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