Prediction of groundwater quality parameter in the Tabriz plain, Iran using soft computing methods

Groundwater is one of the valuable resources of any country. These resources have been taken into account because of the lower pollution potential and the high storage capacity of surface water (Tabarmayeh & Vaezi Hir 2015; Qasemi et al. 2019). Globally 40% of the extracted groundwater is consumed for agricultural purposes (Qasemi et al. 2018a, 2018b; Wagh et al. 2018). In other words, groundwater is one of the main sources of agricultural water supply.

However, in Iran, due to the dry and semi-arid climate conditions, especially in the study area (Tabriz plain located in the eastern part of Lake Urmia), the importance of groundwater conditions becomes more intense. In Tabriz plain, the water needed for agricultural land is provided in addition to surface water resources from groundwater sources (Razaghmanesh et al. 2006). In general, changes in the patterns of land use, climate, urbanization, and population growth and the extensive use of chemical fertilizers and pesticides that penetrate underground in many areas are serious threats to the quality of groundwater, both quantitatively and qualitatively. The groundwater quality is getting worse gradually (Wagh et al. 2016, 2018).

The study of the GFZ German Research Centre for Geosciences, Potsdam (GFZ) states that the surface water table in Iran over the past 15 years has dropped by an average of 1.5 feet (0.4572 m) a year. Groundwater quality is as important as its quantity, and is essential for its usability in various uses. Water quality analysis is an important part of groundwater studies, as the quality of these resources, such as surface water, is constantly changing. However, these changes are much slower than those of surface water (Mahdavi 1995). Modeling groundwater aquifers and consequently predicting the quality of water resources is important in terms of hydrological studies and agricultural management in order to achieve high quality groundwater (Todd & Mays 2005). Hence, the use of intelligent models for the evaluation of groundwater quality parameters has become widespread today and there are many studies in this field.

Nourani et al. (2008) studied temporal and spatial variations of groundwater level in Tabriz plain using the neural network. They used six types of architectures and algorithms to determine the best structure of the neural network to predict the groundwater level. The results of their study showed that the feedback artificial neural network along with the Levenberg–Marquardt (LM) algorithm provide the best forecasts among a variety of other networks.

Asadollahfardi et al. (2011) employed two types of artificial neural networks including recurrent neural network (RNN) and multilayer perceptron (MLP) to predict the total dissolved solids (TDS) in Talkheh Rood. The results showed that the performance of the RNN method was better than MLP in predicting TDS.

Talebizadeh & Moridnejad (2011) used adaptive neuro-fuzzy inference system (ANFIS) and artificial neural network (ANN) models to predict the level of Urmia Lake. The results showed that compared to ANN, the accuracy of the ANFIS model was higher in this regard. Karimi et al. (2012) used gene expression programming (GEP) and ANFIS models to estimate the water level of Urmia Lake. They also concluded that GEP was more efficient compared to the ANFIS model. Zaman Zad Ghvidel & Mntaseri (2014) evaluated the application of different data-based methods in predicting the TDS in Zarinehroud basin. They concluded that GEP, ANFIS with grid partition (ANFIS-GP), ANFIS with sub clustering (ANFIS-SC), and ANN models could be employed to predict the changes in TDS effectively. In addition, among the various methods of artificial intelligence, GEP was the superior method in the prediction of this parameter compared to the other models.

In a study by Kadam et al. (2019) in the Shivganga River in India, ANN and multiple linear regression (MLR) models were used to predict groundwater quality for drinking purposes. In this research, an MLR model was used to evaluate the ANN performance prediction and the results showed that ANN performance was satisfactory.

Taking the improved performances of different modeling techniques into account, this study compared the performance of four soft computing techniques, MLP, ANFIS, support vector machine (SVM), and GEP, for TDS prediction using physicochemical parameters of groundwater collected from the East Azarbaijan Regional Water Organization. It was compared because TDS is an important quality criterion for agricultural water.

Groundwater quality data of the Tabriz plain were obtained from 1,742 samples studied during 2002–2012. The data were collected from Tabriz Regional Water Authority in East Azarbaijan Province. Data were collected twice a year for 10 years, in different piezometers in Tabriz plain. Generally, of the whole data set, 70% (1,220 samples) was used for training and 30% (522 samples) for testing. Excel software was used to prepare the statistical data. Then, four models including MLP, ANFIS, SVM, and GEP were used for modeling TDS. In this regard, the inputs of the model were selected in terms of the highest correlation with TDS (Na, HCO₃, Ca, Mg, and SO₄), respectively. The parameters root mean square error (RMSE), mean absolute error (MAE), and R were used to evaluate and select the best model.

ANN is a model based on the human brain, the idea for which was first proposed in 1940 in a School of Psychiatry. One of the most widely used types of ANN is MLP which is also employed in this study (Kouvhskzadeh & Bahmani 2006). Overall, the MLP neural network structure consists of three important layers, namely, input, hidden, and output layers and in each layer a number of neurons are considered in the network architecture. This model has been successfully used in several studies (Deswal & Pal 2008; Singh et al. 2010; Khatibi et al. 2011; 2013; Ghorbani et al. 2013). In the MLP method, the number of neurons in the input and output layers is determined depending on the number of input and output variables of the system under study, respectively. In order to obtain the parameters of the MLP (number of hidden nodes, the learning rate, and the momentum value), the data set is divided into two parts: training and testing sets. Logistic sigmoid transfer function was applied in the hidden layer. One hundred trials of random initial weights were considered to prevent the local minima problem affecting the MLP method. The network was trained in 1,000 epochs using the LM algorithm with a learning rate of 0.001 and a momentum value of 0.7. Detailed theoretical information about MLP neural networks can be found in Haykin (1998). Figure 1 shows the structure of the multi-layer perceptron with five inputs. MLP models used in this study include an input layer with variable neurons including five input compounds (Na, HCO₃, Ca, Mg, and SO₄) and one hidden layer (with two neurons) and one output layer (TDS).

Fuzzy system and fuzzy logic have an important role in modeling methods. Fuzzy system is a system based on logic ‘If–Then’ rules, that projects the space of input variables on the space of output variables by means of the concept of linguistic variables and fuzzy decision-making process. The combination of fuzzy systems which are based on logical rules with the artificial neural networks which have the ability to extract knowledge from numerical information, have led to the presentation of the neuro-fuzzy inference system (Zadeh 1965; Mohammadi et al. 2018; Takdastan et al. 2018). ANFIS has a five-layer structure of the forwarding neural network and these five layers are: layer 1 (fuzzification layer), layer 2 (rule layer), layer 3 (normalization layer), layer 4 (defuzzification layer), and layer 5 (a single summation neuron) or (output) layer, as shown in Figure 2. The training process in ANFIS is the propagation of error and in order to reach a rapid convergence the combined approach can be used, which is a combination of error propagation with the least squares method (Jang et al. 1997). The ANFIS network used in this study has a membership function of ‘constant’ in the output layer and a membership function of ‘gaussmf’ in the input layer. The membership function of ‘constant’ is constant for each rule. The ANFIS structure is shown in Figure 2. In this figure, (x) and (y) are inputs of the model, (A) and (B) are fuzzy sets, and (f) is the output of the model. In the first layer, all nodes are comparative and the output of layer 1 is the membership grade of the inputs. The output of layer 2 is the input signal coefficient, which is actually equivalent to the (if) rule. The output of layer 3 is normalized to the previous layer. In layer 4, the contribution of each rule is calculated for the final input and determination of the final function. Finally, the output of layer 5 is the total output of the system.

Machine learning methods for the development of classification models are widely applied in numerous fields. The SVM was introduced as a regression tool for the first time by Vapnik (2013). The SVM method is a set of supervised learning algorithms used for classification and regression challenges (Cortes & Vapnik 1995). SVMs are kernel-based algorithms that transform data into a multi-dimensional space and form a hyperplane that enlarges the interval to the nearest data point of any of the input categories (Vapnik et al. 1997). For further information about the SVM method readers are referred to Gunn (1998). The SVM structure is shown in Figure 3. This figure contains input vectors, which are represented by x¹ to xⁿ, the x₁ to x_N parameters are represented as support vectors in the figure, k is the kernel function and the number of support vectors. Finally, the weight is applied to the output of the kernel function and the sum is displayed as output (y).

GEP, based on the principles of genetics, is a symbolic regression technique that automatically solves problems without pre-specified structure of the solution in advance (Babovic 2005). Five major preparatory steps for GEP are: (1) determining the set of terminals, (2) determining the set of functions, (3) determining the fitness measure, (4) determining the parameters for the run, and (5) determining the method for designating a result and the criterion for terminating a run (Koza 1994). The first step in the GEP algorithm is to generate the initial population of solutions. The chromosomes are then evaluated as a tree diagram (ET) with a fitness function to determine the suitability of a solution in the problem domain; if the satisfactory quality of a solution is found, evolution is stopped and the best solution is reported. On the other hand, if the conditions are not stopped, the best solution is kept from the current generation (this means elitist selection) and the rest of the solutions are left to the selective process (Ferreira 2001). For more information about the GEP method, readers are referred to Ghorbani et al. (2012).

The Tabriz plain is located at the eastern part of Urmia Lake in the north-west of Iran. This plain lies between latitude 37° 53′ to 38° 12′ N and longitude 45° 55′ to 46° 45′ E. The elevation of this plain ranges between 1,250 and 3,600 m above sea level. The climate of the area is semi-arid and average annual rainfall is about 230 mm. Figure 4 shows the Tabriz plain located in the eastern part of Lake Urmia in Iran. Figure 4 shows groundwater quality data of observation wells (piezometers) in the aquifer collected by the East Azarbayjan Regional Water Authority of Iran from 2002 to 2012, which were used in this research.

Table 1 provides the basic statistics of the water quality data used in this study. In this table, X_Min, X_Max, X_Mean, S_d, C_v, and C_SX, denote the overall minimum, maximum, mean, standard deviation, coefficient of variation, and skewness values, respectively. This information is provided for the two parts of training and testing.

The more the correlation coefficient (R) is closer to one, the more consistency there is between the estimated values and observational values. The RMSE value represents the root mean square of the error between the measured and estimated values, and the lower the value, the greater the accuracy of the model estimation. The MAE value represents an absolute computation error, which is close to zero, indicating the high accuracy of the model.

MLP models used in this study include an input layer with the neurons changing from 1 to 5 (five input combination) and a hidden layer with two neurons and an output layer. The performance of the MLP models with training and testing data sets are provided in Table 2. The input composition of each scenario was obtained using the correlation between the physicochemical parameters and the TDS. Accordingly Na, HCO₃, Ca, Mg, and SO₄ parameters had the highest correlation with TDS, respectively, which were used in the scenarios. The model is represented by one input as MLP1 and the model is represented by two inputs as MLP2. Finally, the model with five inputs is shown as MLP5.

According to the table, with the increasing number of inputs in the model, the results will be better (according to the evaluation criteria). This result is consistent with the results of Khashei et al., who stated that the MLP model increases with increasing the number of input parameters, thus the accuracy of the model increases (Khashei et al. 2011; Khashei & Sarbazi 2015) so that the fourth and fifth scenarios are very close together. Therefore, the MAE value is better than the fifth scenario in both the test and training phases in the fourth scenario. However, considering the high R (in the test and training phase) and the lower RMSE (in the testing phase) in the fifth scenario, the MLP model was accepted as the best model in the fourth scenario (R = 0.998, RMSE = 67.67, and MAE = 17.34). According to Figures 5 and 6, we see that there is a very good overlap between the data obtained from the model and observational data.

In Figure 6, R = 0.9985 indicates that these conditions are optimal. Thus, it can be stated that the fifth scenario (MLP5) predicts the high accuracy of TDS data.

The ANFIS method is implemented using the Fuzzy Logic Toolbox (MATLAB) for all short-term flow forecasting in this study. The number of membership functions are the same and equal to 4 for all inputs and membership function types are triangular-shaped, respectively. For performance of the ANFIS models, 70% of the data were for training and 30% for testing, as provided in Table 3. This table indicates that according to the selected evaluation criteria in this study, the results of the models with three inputs, four inputs, and five inputs are very similar. According to table information, the results of the model will be improved as the number of inputs increases. Therefore, the results for the ANFIS model with five inputs are better than other inputs; also the ANFIS model with five inputs, like the MLP5 model, has the best performance. According to the table, and at best, the values of the parameters are R⁼0.997, RMSE = 108.63, and MAE = 27.77. The output of the data obtained with the model and observational data are shown in Figure 7. As shown in the figure, there is an overlap between estimated data and observational data. In Figure 8, the value of this overlap is shown as a correlation coefficient (R = 0.9977), which expresses the high accuracy of the model ANFIS5 for predicting computational TDS.

Here, the SVM method was used for predicting the TDS of groundwater dominated by Urmia Lake salt water. The main benefit is to avoid attributes in greater numeric ranges dominating those in smaller numeric ranges. Another benefit is to avoid numerical difficulties during the computing. For the implementation of the SVM models, as with the previous models, 70% of the data were selected for training and 30% testing. The preparation of the model with different inputs (1 to 5) is shown in Table 4. This table indicates that the model results with two inputs (SVM2), four inputs (SVM4), and five inputs (SVM5) are very close together. However, in the end, according to the best assessment criteria in the fifth scenario, test value (R = 0.953, RMSE = 274.56, and MAE = 13.83) and training value (R = 0.958, RMSE = 471.78, and MAE = 7.25), the top model, SVM5 was selected. The results of the best model, SVM with five inputs (SVM5), are shown in Figures 9 and 10, illustrating that the predicted data are obtained with higher accuracy compared to the observational data, and clearly evident from the high (R = 0.9533) in Figure 10. Therefore, the SVM model, like the ANFIS and MLP models with five inputs (fifth scenario), has the best performance for predicting TDS.

Various GEP models have been developed using input combinations similar to MLP, ANFIS, and SVM models. The performance of the GEP models with training and testing data sets are provided in Table 5. Unlike the MLP, ANFIS, and SVM models, where the fifth scenario was the best option, here, according to the evaluation criteria, the GEP model with four inputs has the best performance. This table shows that the GEP model with four inputs has a better performance than the other input combinations in terms of R, RMSE, and MAE values. Prediction accuracy of the GEP4 model is much better than other GEP models used with this data set in terms of R (0.998), RMSE (58.93), and MAE (5.21) values. Figure 11 shows the observational data and the data obtained from the output of the GEP4 model. It is clear that there is a high correlation between estimated data and observational data in Figure 12, with respect to the value of R (0.9992). This high correlation is proven. Therefore, according to the results obtained from the models, and considering the numbers and variables of the evaluation parameters of the models (R, RMSE, MAE), it was found that the GEP model was selected as the best model for predicting qualitative parameter (TDS) compared to other models of this study (MLP, ANFIS, SVM). This result is consistent with the studies of Zaman Zad Ghvidel & Mntaseri (2014) and Karimi et al. (2012). Given that the GEP model offers a mathematical equation, the equation of the optimal model in MATLAB format is shown in Figure 13. Also, Figure 14 shows the expression tree of the optimal GEP model (GEP4) for TDS estimation. The tree structure helps in each stage that the initial population be expressed in a simple linear structure, and all the changes are made to simple structures. As a result, there is no need for relatively complex structures to be expanded at each stage. In Figures 13 and 14, the values of C are related to the coefficients of gene expression planning and the values of d are related to the input parameter.

Soft computing approaches can provide relatively accurate predictions for water quality parameters; however, these intelligent methods rely on data sets that take a long time in operation. In the current study, the applicability of MLP, ANFIS, SVM, and GEP models in predicting TDS values has been evaluated. Groundwater samples of Tabriz plain in northwest Iran were tested for physicochemical parameters using different models. Regarding the correlation coefficient of physicochemical parameters to TDS, the inputs of the model were selected according to the selected priority (Na, HCO₃, Ca, Mg, and SO₄) and in the scenarios 1 to 5. Finally, they were analyzed by the criteria of the evaluation (R, RMSE, MAE). The MLP model for scenarios 4 and 5, according to the RMSE, MAE, and R assessment criteria, had roughly similar results: MLP4 for test (R = 0.997, RMSE = 83.81, and MAE = 6.52) and MLP5 for test (R = 0.998, RMSE = 67.67, and MAE = 17.34). Therefore, considering the greater correlation coefficient (R) and lesser RMSE, the fifth scenario was chosen as a better scenario. The MLP, ANFIS, and SVM models had the best performance with five inputs, but the GEP model with four inputs was chosen as the best model. Also, according to the results, the MLP model was better than the ANFIS and SVM models according to (R = 0.998, RMSE = 67.67, and MAE = 17.34) evaluation criteria. On the other hand, the MLP model with five inputs (MLP5) and the GEP model with four inputs (GEP4) had similar results, but in the end, the GEP model was selected as the best model according to better conditions (R = 0.998, RMSE = 58.93, MAE = 5.21). Also, this method can provide mathematical modeling that is very effective for future predictions.

The authors want to thank the authorities of Islamic Azad University for their comprehensive support for this study. The authors of this article declare that they have no conflict of interests.