Listen

This paper presents to study the performance of machine learning techniques consisting of multivariate adaptive regression spline (MARS), feed forward neural network-back propagation (FFNN-BP), and decision tree regression (DTR) for estimating the physico-chemical properties of groundwater in the coastal plain area in Vinh Linh and Gio Linh districts of Quang Tri province of Vietnam. With 290 groundwater samples collected in two districts, this study has identified three main elements CO2, Ca, CaCO3 for simulation. Quantitative analysis results have shown that these three components are such as CaCO3 with from 0 to 25.8 mg/l, Ca from 0 to 87.55 mg/l and CO2 from 0 to 12 mg/l. In the present examination, groundwater quality index (GQI) values and their representative categories have been referred by the Vietnam Groundwater Standard (QCVN01). Furthermore, the statistical accuracy parameters were used to compare among models. To deploy FFNN-BP and DTR, different types of transfer and kernel functions were tested, respectively. Determining the results of MARS, FFNN-BP and DTR showed that three models have suitable carrying out for forecasting water quality components. Comparison of outcomes of MARS model with the FFNN-BP and DTR models indicated that this model has good performance for forecasting the elements of water quality, its level of accuracy was slightly more than the other. To assess the accurate values of the models according to the measurement parameters for training phase illustrated that the order of the models was MARS to give the best result, followed by DTR and finally FFNN-BP, respectively.

Listen

  • Machine learning methods are used for spatial modeling of physico-chemical properties of groundwater.

  • MARS performances suitable precision compared to the DTR and FFNN-BP models.

  • Total CaCO3 value in the experiment samples adapted the regular limit of QCVN01 with ‘Excellent’ point.

  • The quality of water parameters (i.e., CaCO3, Ca, and CO2) of the coastal plain area was predicted.

  • The study results have shown that the water quality in these two districts is usable for humans, livestock, and agriculture activities.

groundwater, machine learning, physico-chemical properties, prediction
Listen

The presence of contaminants in natural freshwater is considered one of the most crucial environmental problems in many areas of developing countries, where several communities are hardly approaching a potable water supply (Organisation mondiale de la santé, Światowa Organizacja Zdrowia, World Health Organization, & World Health Organisation Staff 2004; Giang et al. 2021). Low-income communities, which lean on untreated surface and groundwater supplies for domestic and agricultural purposes are the most affected by poor water quality (Ayoko et al. 2007). Unfortunately, they also do not have adequate tools to monitor quality of water regularly (Resh 2008; Omarova et al. 2018; Najafzadeh & Niazmardi 2021). Thus, they are increasingly expected to obtain reliable assessments of quality of water, which can be used (Bonansea et al. 2015).

Climate change leads to seawater intrusion affecting groundwater resources of coastal cities (Kumar 2012; Alfarrah & Walraevens 2018). Urbanization and industrialization have caused uncontrolled over-exploitation and depletion of groundwater consequently (Kanwal et al. 2015; Sunardi et al. 2021). Furthermore, untreated wastewater from residential areas and industrial zones has seeped into the ground leading to an increasing amount and content of chemical elements in groundwater (Mukate et al. 2018; Khan et al. 2020a).

The chemical element of groundwater is considered a standard of measurement to show the capable level of groundwater for plenty of targets such as human and animal drinking, agricultural, and industrial activities. It has been shown in practice that the uses of groundwater sources require different standard indicators to distinct water quality circumstances (Loaiciga et al. 1992). The groundwater quality concept is an integrative index composed of chemical, physical, and biological features which maintain expected groundwater utilizations. Hence, groundwater is divided by composition as groundwater quality index (GQI) for the management and consumption of groundwater resources (Najafzadeh et al. 2021a).

To evaluate the quality of water for drinking and agricultural irrigation, several variables are routinely monitored. This process makes a big database, but it can be time-consuming for data acquisition while the accurate rendering of the multivariate data may be challenging.

With regard to use machine learning for forecasting physico-chemical parameters in water, using artificial neural network (ANN)-estimated river water quality components (Niroobakhsh et al. 2012; Emamgholizadeh et al. 2014; Najah et al. 2014; Raheli et al. 2017; Haghiabi et al. 2018; İlhan et al. 2021; Najafzadeh et al. 2021b); employing multivariate adaptive regressive splines (MARS) to predict physico-chemicals in water (Haghiabi 2016; Bhatt et al. 2017; Ahmadi et al. 2019; Esmaeilbeiki et al. 2020); deploying decision tree regression (DTR) to forecast quality of water (Liao & Sun 2010; He et al. 2012; Jaloree et al. 2014; Chandanapalli et al. 2018; Gakii & Jepkoech 2019; Jalal & Ezzedine 2020; Lu & Ma 2020). Furthermore. MARS, feed forward neural network-back propagation (FFNN-BP), and DTR models also belong to nonparametric learning, and the model is used in those areas (Bengio et al. 2010; Al Iqbal et al. 2012; Genuer et al. 2017; Khaldi et al. 2019; Kohler et al. 2019; Yurochkin et al. 2019; Antoniadis et al. 2020; Devianto et al. 2020; Khan et al. 2020b; Zheng et al. 2020; Amiri-Ardakani & Najafzadeh 2021). Najafzadeh & Ghaemi (2019) implemented the LS-SVM and MARS models to estimate BOD5 and COD parameters through 200 samples collected from Karoun River, in the southwest of Iran. The result showed that the MARS model has proved precise approximations compared with real data. Saghebian et al. (2014) applied a decision tree model to classify groundwater quality in Ardebil, Iran. Research results have proved that this model can be acceptable range of criteria for quality classification of groundwater. Khan et al. (2021) used FFNN-BP model to estimate Escherichia coli in groundwater with 1,301 groundwater samples were obtained from 348 villages and cities in from 2016 to 2019 in Rajasthan state, India. Consequently, deploying the model based on Grover's algorithm was more efficient in forecasting all patterns in the calculated E. coli in groundwater. Najafzadeh et al. (2021a) studied the groundwater quality of the Rafsanjan Plain of Iran, quantifying it using artificial intelligence (AI) to assess GQI values for 15 years. The results of the groundwater quality prediction analysis of the MARS model with RMSE = 2.444 and SI = 0.0304. In addition, this result was also compared with the World Health Organization groundwater standard which also showed that the entire area of Rafsanjan lacks water quality at the ‘Excellent’ level with a high probability. The chance for ‘Good’ water quality varies from 1% (at GQI = 50 worst cases) to 55% (at GQI = 100 best cases).

Groundwater quality prediction work has some errors for various reasons such as the quality of collected groundwater samples, measurement of variability and the subjective opinion of groundwater sample analysts, and other random parameters related to groundwater quality prediction that have not been studied yet. Therefore, the problem of assessing reliability for implicit quality classifications. In addition, using analytical methods is also subject to the bias of environmentalists, geologists, and experts.

This paper presents the prediction of the physico-chemical properties of groundwater using FFNN-BP, DTR, and MARS models. The input vectors used in the models are leaned on 290 samples that were collected from 290 wells of households in coastal plain area in Gio Linh and Vinh Linh districts of Quang Tri province. With the support of the collected data, the GQI values were analyzed based on the Vietnam Groundwater Standard (QCVN01) and their relevance for proposed use. After that, highlight comparison among three models that base on the results of statistical accuracy parameters such as mean (M), bias (bias is shown by mean error), root mean square error (RMSE), mean absolute error (MAE), standard deviation (St Dev), pearson correlation coefficient (R), skewness coefficient (Skew), minimum (Min), maximum (Max), scatter index (SI), and Nash–Sutcliffe efficiency (NSE). Finally, the collection of results of these three models may show the working efficiency of the models for predicting the quality of water.

The structure of the paper is organized as follows. Section 1 gives the paper's introduction. Section 2 presents study area, the MARS, FFNN-BP, and DTR models and explains them clearly for understanding use throughout this paper. Section 3 describes the study results. Finally, Section 4, and Section 5 introduce the discussions and conclusions.

Listen

Study area

Listen
The place of study is about 150 km2 and covers the Gio Linh and Vinh Linh coastal plain of Quang Tri province of Vietnam. It is surrounded by Quang Binh province in the north, Thach Han River in the south, 50–150 m high hills in the west, and the East Sea (see Figure 1). The coastal plain is relatively flat with an elevation between 0 and 5 m except for coastal sand-dunes at 11–22 m high, which provide a natural embankment system for seawater protection (Krutwagen 2007). During the dry season (from June to August), the saltwater pervades (i.e. where total mineralization of water M = 1 g/l) and is often inspected at about 30 and 35 km from the main tributaries of Ben Hai, Hieu, and Thach Han Rivers from the estuary (Tam et al. 2014). Thus, groundwater from dug wells and shallow wells scattered in the coastal is the main water source for drinking and domestic use for residents.
Figure 1
The location for samples collection.
View largeDownload slide

The location for samples collection.

Figure 1
The location for samples collection.
View largeDownload slide

The location for samples collection.

Close modal

MARS

Listen
The MARS model is a novel approach in soft computing, and it is a nonparametric regression model, introduced by Friedman (Friedman et al. 2010). MARS seems like a method for a fitted relationship between independent and dependent variables in each desired phenomenon. MARS supports techniques for modelling systems with high accuracy, which is based on a dataset (Sekulic & Kowalski 1992; Steinberg 2001; Gutiérrez et al. 2009). The MARS algorithm feature is the procedure of the backward and forwards stepwise, at the same time may explain and control the complex nonlinear mapping between the inputs and output variables. MARS model highlights input variables that have a note worthy effect on the output variables. The general form of MARS is described as below:
(1)
where, y is output variables, is constant value, M is the number of functions, is basis function and is the corresponding coefficient of Furthermore, shows information about the relationship between input and output variables, and it is described as below:
(2)
where h is the basis function, x is the input variables, and C is the threshold value of the independent (input) variables of x. It is notable that C is called ‘knots’ or ‘hinges’.
The function of backward stepwise function relates to removing basis functions one at a time until the criterion of ‘lack of fit’ is a minimum. In the deletion of backward stepwise, the last important basic functions are destroyed one at a time. The lack of used fitting measurement is based on generalized cross-validation (GCV) (Attoh-Okine et al. 2009):
(3)
where are the complexity function (Friedman et al. 2010). The GCV criterion is considered the average of residual error multiplied by a penalty to modify for the variability associated with more parameters estimation in the model (LeBlanc & Tibshirani 1994).

Feed forward neural network-back propagation

Listen
FFNN-BP model is a member of neural network method (Khoshhal & Mokarram 2012). It may simulate arbitrarily complex nonlinear processes for any systems in terms of inputs and outputs. A FFNN-BP structure in Figure 2 demonstrates a three-layer neural network consisting of inputs layer, hidden layer (layers) and outputs layer (Ramchoun et al. 2016; Ashiquzzaman & Tushar 2017). The physico-chemical properties groundwater presents neural network that have trained for its FFNN-BP regression. The input layer has 290 input nodes from e1 to e290, one neuron of the output layer has deputized the values of physico-chemical properties groundwater. There are four hidden layers include the first hidden layer contains neurons from H11 to H1100, the second layer is from H21 to H2100, the third layer is from H31 to H3100, and the last one is from H41 to H4100. Each neuron of the hidden and output layers get a corresponding weight and bias, as ⁠, and ⁠, are the weight and bias to represent for neurons of H11 and H12, so on. The values of weight and bias can be assigned progressively and corrected during the training process in order to compare predicted outputs with known outputs. As networks are often trained using a backpropagation algorithm (Ashiquzzaman & Tushar 2017). Each neuron of the hidden layers attains the output from all neurons of the previous layers and converts these values with a weighted linear sum into the output layer. The output layer receives the values from the last hidden layer. The ReLU function is deployed as the activation function for the hidden layers. Adam method is stochastic optimizations to the solver of weight optimization.
Figure 2
Structure of FFNN-BP network for physico-chemical properties groundwater prediction.
View largeDownload slide

Structure of FFNN-BP network for physico-chemical properties groundwater prediction.

Figure 2
Structure of FFNN-BP network for physico-chemical properties groundwater prediction.
View largeDownload slide

Structure of FFNN-BP network for physico-chemical properties groundwater prediction.

Close modal

DTR

Listen
A decision tree is a data structure that includes an arbitrary number of nodes and branches at each node (Pekel 2020). The values of the input variable(s) consider a particular function in the training stage (Loh 2011). The stimulant of the decision tree is an algorithm that generates a decision tree from given instances. The structure of decision tree regression is as below, assuming X = X1, X2, …, Xmn, mn are estimator variables, a total number of estimator variables, respectively. At the same time, n and Y = Y1, Y2, …, Yn describe the number of observations and a goal variable that doing continuous values, respectively. In addition, vf, th, t, and γ (γ = (vf, tht)) are a characteristic of a variable, a value of threshold, a node, and a candidate split, respectively. Ql is realized by splitting the data into γ. and are calculated in Equations (4) and (5) declare the left and right sides in the decision tree of Q.
(4)
(5)
Let calling nl and nr are the number of a sampling of left and right sides; I is a function of impurity. Equation (6) indicates how the function of impurity is computed. The function of impurity is minimized by pondering Q and γ indicators as:
(6)

Performance metrics

Listen
Forecasting results are based on the calculation and comparison of the actual values to the forecasted values. These metrics of the accuracy measurement parameters include MAE, RMSE, NSE, bias, SI. Furthermore, the error metrics are defined as follows (Yang & Yang 2005; Touzani et al. 2018; Kardani et al. 2020):
(7)
(8)
(9)
(10)
(11)
(12)
where are the estimated value and observed value in the period time t, and n is the number of the observed values in the testing data. are mean of the observed value. The should be approaching 1 to indicate strong model performance, and the bias, MAE, and RMSE should be as close to zero as possible.

Data collection

Listen
Study data includes 290 groundwater samples obtained at wells of households in coastal plain area of Vinh Linh and Gio Linh districts. The predominant chemical compositions in these samples consisted of three main ingredients as calcium carbonate (CaCO3) calcium (Ca), and carbon dioxide (CO2). In addition, there were some other physico-chemical components (as ammonia, magnesium, and iron oxide), but their contents were not significant in these samples. Three input variables include Ca, CO2, and CaCO3, which was collected from 290 wells of two districts’ households. The statistical characteristic results are also pointed out in Table 1. The range of the following characteristics was computed from the observation: the mean, min, and max values, St Dev, skew. The mean and standard deviation of the CaCO3, Ca and CO2 were 1.30 mg/l and 4.23 mg/l, 6.05 mg/l and 16.1 mg/l, 0.79 mg/l and 2.42 mg/l, respectively. The skewness for a normal distribution is zero, and any symmetric data should have a skewness near zero. Negative values for the skewness indicate skewed left data, and positive values for the skewness indicate skewed right data (Sahu et al. 2003; Brys et al. 2004). Hence, the skew of data fluctuating from 1.56 mg/l to 2.06 mg/l could be considered acceptable for prediction through these models. The input data patterns of 290 items were randomly selected with two parts. The first part was used for the training phase, which contained about 70% of the entire data. The second part was used for the test phase, which contained about the remaining 30%. In addition, the methodology of this study is described by the diagram in Figure 3. The process was summarized by experimental stages as below. Firstly, the collected dataset is preprocessed and tested statistical procedure, and the data is also divided into training phase and testing phase. Secondly, the FFBB-PB, MARS, and DTR models are employed based on the training samples, and to acquire the best network parameters. Finally, the performances of the algorithms are compared by using metrics of the accuracy parameters, and looking for the most suitable forecasting model is found for the study.
Table 1

Statistical characteristics of physico-chemical components data

ItemSt DevMeanMinMaxSkew
CaCO3 4.23 1.30 0 25.80 2.06 
Ca 16.1 6.05 0 87.55 1.57 
CO2 2.42 0.79 0 12 1.56 
ItemSt DevMeanMinMaxSkew
CaCO3 4.23 1.30 0 25.80 2.06 
Ca 16.1 6.05 0 87.55 1.57 
CO2 2.42 0.79 0 12 1.56 

Unit: mg/l.

View Large
Figure 3
Flowchart of the experimental steps conducted in this study.
View largeDownload slide

Flowchart of the experimental steps conducted in this study.

Figure 3
Flowchart of the experimental steps conducted in this study.
View largeDownload slide

Flowchart of the experimental steps conducted in this study.

Close modal
Listen

The output function of MARS is presented as below:

MARS = 7.907–0.249F1–0.129F2, where F1 = max(0, Ca − 55.79), F2 = max(0, 55.79 − Ca).

Fi is the basis function. F1 may be explained as the maximum value of 0 and Ca − 55.79. The minus sign ahead of the maximum value is equivalent to a minimum value. In addition, the MARS analysis indicates that the most important is Ca. Furthermore, the output function for FFNN-BP and DTR do not occur.

The data in Figure 4(a)–4(c) shows the relationship between the three variables. The content of CO2 and Ca increase lead to the content of CaCO3 increase. The FFNN-BP model makes a forecasting form that resembles a cone shape. In the meantime, the DTR and MARS charts look like the image of papers with some folds. Through these three images, it is hard to judge which model gives the best estimating. Hence, the values of performance metrics of the three models are presented in Table 2, Figure 5. The NSE of the three models for the training and testing phases are from 0.89 to 0.95 and are closer to 1. The MAE, RMSE, and bias values are also from −0.12 mg/l to −0.09 mg/l and are closer to 0. In addition, the values of the SI statistical indicator are simulated fluctuation from 0.21 mg/l to 2.23 mg/l. These show that the forecast results are very consistent compared with the actual data. As for the experimental results for each specific model, it indicates that the MARS model for training phase with NSE, MAE, RMSE, bias, and SI values are 0.95, 0.14 mg/l, 0.24 mg/l, 0.00 mg/l, and 0.26 mg/l respectively, and these properties are better than DTR and FFNN-BP models. Regarding the testing phase results, the highest accuracy for forecast is MARS model with NSE = 0.95, the second-highest is FFNN-BP model with NSE = 0.91, and the lowest is DTR model with NSE = 0.89.

The Taylor charts check the performance of estimated and actual values based on the standard deviation and Pearson Correlation Coefficient (Qin & Xiao 2018), which are contained simultaneously in assessing the respective models (Taylor 2001; Ghorbani et al. 2018). The standard deviation and CC between the actual and predicted datasets for the models are present in the Taylor diagram, and it also can be seen overall consistency between observed and estimated values when the CC value is approaching up to 1, as pointed in Figure 6. This can be considered for the MARS model with CCtraining phase = 0.94, CCtesting phase = 0.93, FFNN-BP model with CCtraining phase = 0.91, CCtesting phase = 0.92, and DTR model with CCtraining phase = 0.93, CCtesting phase = 0.91. The large number of correlation coefficients indicate that there is a strong relationship. The Taylor plot also shows that these models are optimal with the highest accuracy (Taylor 2001). In other words, if the standard deviation of the predicted value of the higher standard deviation of the observed value, it will lead to an over estimation and vice versa (Abba et al. 2020). Furthermore, the GCV indicator of MARS brings about equilibrium between flexibility and generalization ability of the function of MARS model (Deo et al. 2016).
Table 2

Accuracy parameters for physico-chemical components prediction

ParameterDTR
FFNN-BP
MARS
TestingTrainingTestingTrainingTestingTraining
MAE (mg/l) 0.25 0.25 0.50 0.46 0.21 0.14 
RMSE (mg/l) 0.34 0.33 0.99 0.90 0.41 0.24 
Bias (mg/l) −0.09 −0.19 −0.12 0.01 −0.04 0.00 
SI (mg/l) 3.23 3.10 1.53 1.19 0.21 0.26 
R 0.91 0.93 0.92 0.91 0.93 0.94 
NSE 0.89 0.94 0.91 0.90 0.95 0.95 
GCV (mg/l)     0.14 0.14 
ParameterDTR
FFNN-BP
MARS
TestingTrainingTestingTrainingTestingTraining
MAE (mg/l) 0.25 0.25 0.50 0.46 0.21 0.14 
RMSE (mg/l) 0.34 0.33 0.99 0.90 0.41 0.24 
Bias (mg/l) −0.09 −0.19 −0.12 0.01 −0.04 0.00 
SI (mg/l) 3.23 3.10 1.53 1.19 0.21 0.26 
R 0.91 0.93 0.92 0.91 0.93 0.94 
NSE 0.89 0.94 0.91 0.90 0.95 0.95 
GCV (mg/l)     0.14 0.14 
View Large
Figure 4
Physico-chemical properties prediction with (a) MARS model, (b) FFNN-BP model, and (c) DTR model (Unit: mg/l).
View largeDownload slide

Physico-chemical properties prediction with (a) MARS model, (b) FFNN-BP model, and (c) DTR model (Unit: mg/l).

Figure 4
Physico-chemical properties prediction with (a) MARS model, (b) FFNN-BP model, and (c) DTR model (Unit: mg/l).
View largeDownload slide

Physico-chemical properties prediction with (a) MARS model, (b) FFNN-BP model, and (c) DTR model (Unit: mg/l).

Close modal
Figure 5
The best performance indicators for CaCO3 prediction (a) MARS training model, (b) FFNN-BP training model, (c) DRT training model.
View largeDownload slide

The best performance indicators for CaCO3 prediction (a) MARS training model, (b) FFNN-BP training model, (c) DRT training model.

Figure 5
The best performance indicators for CaCO3 prediction (a) MARS training model, (b) FFNN-BP training model, (c) DRT training model.
View largeDownload slide

The best performance indicators for CaCO3 prediction (a) MARS training model, (b) FFNN-BP training model, (c) DRT training model.

Close modal
Figure 6
The best performance indicators for CaCO3 prediction for Training, and Testing.
View largeDownload slide

The best performance indicators for CaCO3 prediction for Training, and Testing.

Figure 6
The best performance indicators for CaCO3 prediction for Training, and Testing.
View largeDownload slide

The best performance indicators for CaCO3 prediction for Training, and Testing.

Close modal
Listen

Invasive seawater, untreated wastewater from residential areas and industrial zones, and over-exploitation of groundwater have been seriously affecting the quality and quantity of the underground water system. Therefore, a water quality assessment is a regular and continuous work to help people and authorities show solutions to treat groundwater to serve daily life. Hence, this paper described a comparative study and analysis of MARS, FFNN-BP, and DTR models in estimating physico-chemical properties of groundwater. The different circumstances, influential factors, and indicators have been observed for the experimentation. The following key findings are as the predictive errors in the case of the models decreased if the testing set decreased; MARS was the highlight in comparison to other models. Furthermore, to compare RMSE and MAE using MARS model of this study result with the study result of Najafzadeh et al. (2021a) about the groundwater quality indicate that their RMSE = 0.55 mg/l, and MAE = 0.00 mg/l are equivalent this study with RMSE = 0.41 mg/l, MAE = 0.21 mg/l.

According to QCVN01, the groundwater index of CaCO3 is from 0 to 300 mg/l (National technical regulation on domestic water quality of Vietnam, QCVN 01: 2009/BYT 2021). The indicators in Table 3 show the range of water quality categorization based on the quality index of weight groundwater for human consumption. However, the total CaCO3 value in the experiment samples ranged from 0 to 25.8 mg/l, and the mean value was 1.3 mg/l, this point illustrated that adapted the regular limit of QCVN01 with ‘Excellent’. These points indicate that the water quality in these two districts is usable for humans, livestock, and agriculture activities. Rapid urbanization, industrialization, and climate change will negatively affect the groundwater resource in the areas. Therefore, households need to build a water purification system for drinking; at the same time, the local government needs to supply a clean water system so that people have clean water for daily living. In addition, local authorities and households should install early warning sensors for changes in chemical content in groundwater at some wells. The work periodically checks and advises on pollution levels of groundwater.

Table 3

Water quality categorization based on the quality index of weight groundwater (GWQI)

GWQI< 5050 − 100100–200200–300> 300
Quality categorization Excellent Good Bad Very bad Unsuitable for drinking 
GWQI< 5050 − 100100–200200–300> 300
Quality categorization Excellent Good Bad Very bad Unsuitable for drinking 

Unit: mg/l.

View Large

The groundwater samples for the study were mainly collected at the beginning of the dry season, so the hydrodynamic coefficient is dissimilar from the rainy season. The groundwater analysis equipment had an effect on the study because the equipment did not detect any more chemicals in the water that affect public health in these two areas. However, the results of this study also contribute to supporting local authorities to have appropriate solutions to help households use clean water.

Listen

Assessing and estimating water quality is a difficult and complex task, this result will give warnings to users and authorities to have appropriate treatment solutions. Hence, this study has deployed the MARS, DTR, and FFNN-BP predicted physico-chemical properties groundwater in the coastal plain of Vinh Linh and Gio Linh, which is located in the north middle of Vietnam. For phases of training and testing carried out in the models, the observed data consisting of CO2 and Ca was used as inputs, while CaCO3 was used as output. The stimulated results pointed out that the three models have a high suitable presentation for forecasting water quality components. The best performance was related to the MARS. The results of DTR and FFNN-BP also showed that their accuracy is a suitable presentation for practical purposes. Furthermore, the conducting of a comparison of three models showed that the outcomes of MARS and DTR models were slightly more reliable in comparing with FFNN-BP. The qualitative description of the GWQI found that the whole region of Vinh Linh and Gio Linh districts gained ‘Excellent’ water quality with 100% cases. At the same time, the study results have shown that the water quality in these two districts is usable for humans, livestock, and agriculture.

In addition, this study demostrated that machine learning models play a key role in the decision-making progress for carring out an effects of climate changes, urbanization, and industrialization on quality of groundwater. Another possible future work is to enhance quality of groundwater analysis equipment may finding other elements in the groundwater, and samples should be collected evenly throughout the seasons of the year.

Listen

We would like to thank Quang Tri province of Departments of Natural Resources and Environment, and Science and Technology, and Vinh Linh and Gio Linh residence. We also thanks the project of Hue University ‘Assessing the corrosion capacity of groundwater for foundation concrete structures in the northern coastal plains of Quang Tri province’, the code DHH2013-01-151 has supported the collection of field data in Quang Tri province.

Listen

Conceptualization, Discussion, Writing, Material, Methods, Review, and Editing: Nguyen Hong Giang, Tran Dinh Hieu; Data collection, Writing: Hoang Ngo Tu Do; Methods: Thinh-Tien Nguyen. All authors have read and agreed to the published version of the manuscript.

Listen

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

Listen

The authors declare there is no conflict.

Abba
S. I.
,
Pham
Q. B.
,
Usman
A. G.
,
Linh
N. T. T.
,
Aliyu
D. S.
,
Nguyen
Q.
 &
Bach
Q. V.
2020
Emerging evolutionary algorithm integrated with kernel principal component analysis for modeling the performance of a water treatment plant
.
Journal of Water Process Engineering
33
,
101081
.
Google Scholar
Crossref
Search ADS
 
Ahmadi
M. H.
,
Mohseni-Gharyehsafa
B.
,
Farzaneh-Gord
M.
,
Jilte
R. D.
,
Kumar
R.
 &
Chau
K. W.
2019
Applicability of connectionist methods to predict dynamic viscosity of silver/water nanofluid by using ANN-MLP, MARS and MPR algorithms
.
Engineering Applications of Computational Fluid Mechanics
13
(
1
),
220
–
228
.
Google Scholar
Crossref
Search ADS
 
Al Iqbal
M. R.
,
Rahman
S.
,
Nabil
S. I.
 &
Chowdhury
I. U. A.
2012
Knowledge based decision tree construction with feature importance domain knowledge
. In:
2012 7th International Conference on Electrical and Computer Engineering
(Pran Kanai Saha, ed.). IEEE
, Dhaka, Bangladesh, pp.
659
–
662
.
Google Scholar
Crossref
Search ADS
 
Amiri-Ardakani
Y.
 &
Najafzadeh
M.
2021
Pipe break rate assessment while considering physical and operational factors: a methodology based on global positioning system and data driven techniques
. Water Resources Management 35(11), 3703–3720.
Google Scholar
 
Antoniadis
A.
,
Lambert-Lacroix
S.
 &
Poggi
J. M.
2020
Random forests for global sensitivity analysis: a selective review
.
Reliability Engineering & System Safety
206,
107312
.
Google Scholar
 
Ashiquzzaman
A.
 &
Tushar
A. K.
2017
Handwritten Arabic numeral recognition using deep learning neural networks
. In:
2017 IEEE International Conference on Imaging, Vision & Pattern Recognition (icIVPR)
(Md. Atiqur Rahman Ahad, ed.).
IEEE
, Dhaka, Bangladesh, pp.
1
–
4
.
Google Scholar
Crossref
Search ADS
 
Attoh-Okine
N. O.
,
Cooger
K.
 &
Mensah
S.
2009
Multivariate adaptive regression (MARS) and hinged hyperplanes (HHP) for doweled pavement performance modeling
.
Construction and Building Materials
23
(
9
),
3020
–
3023
.
Google Scholar
Crossref
Search ADS
 
Ayoko
G. A.
,
Singh
K.
,
Balerea
S.
 &
Kokot
S.
2007
Exploratory multivariate modeling and prediction of the physico-chemical properties of surface water and groundwater
.
Journal of Hydrology
336
(
1–2
),
115
–
124
.
Google Scholar
 
Bengio
Y.
,
Delalleau
O.
 &
Simard
C.
2010
Decision trees do not generalize to new variations
.
Computational Intelligence
26
(
4
),
449
–
467
.
Google Scholar
Crossref
Search ADS
 
Bhatt
A. H.
,
Karanjekar
R. V.
,
Altouqi
S.
,
Sattler
M. L.
,
Hossain
M. S.
 &
Chen
V. P.
2017
Estimating landfill leachate BOD and COD based on rainfall, ambient temperature, and waste composition: exploration of a MARS statistical approach
.
Environmental Technology & Innovation
8
,
1
–
16
.
Google Scholar
Crossref
Search ADS
 
Brys
G.
,
Hubert
M.
 &
Struyf
A.
2004
A robust measure of skewness
.
Journal of Computational and Graphical Statistics
13
(
4
),
996
–
1017
.
Google Scholar
Crossref
Search ADS
 
Chandanapalli
S. B.
,
Reddy
E. S.
 &
Lakshmi
D. R.
2018
DFTDT: distributed functional tangent decision tree for aqua status prediction in wireless sensor networks
.
International Journal of Machine Learning and Cybernetics
9
(
9
),
1419
–
1434
.
Google Scholar
Crossref
Search ADS
 
Devianto
D.
,
Permathasari
P.
,
Yollanda
M.
 &
Ahmad
A. W.
2020
The model of artificial neural network and nonparametric MARS regression for Indonesian composite index
. In
IOP Conference Series: Materials Science and Engineering
, Vol.
846
, No.
1
, p.
012007
,
IOP Publishing
.
Google Scholar
Crossref
Search ADS
 
Emamgholizadeh
S.
,
Kashi
H.
,
Marofpoor
I.
 &
Zalaghi
E.
2014
Prediction of water quality parameters of Karoon River (Iran) by artificial intelligence-based models
.
International Journal of Environmental Science and Technology
11
(
3
),
645
–
656
.
Google Scholar
Crossref
Search ADS
 
Esmaeilbeiki
F.
,
Nikpour
M. R.
,
Singh
V. K.
,
Kisi
O.
,
Sihag
P.
 &
Sanikhani
H.
2020
Exploring the application of soft computing techniques for spatial evaluation of groundwater quality variables
.
Journal of Cleaner Production
276
,
124206
.
Google Scholar
Crossref
Search ADS
 
Friedman
J.
,
Hastie
T.
 &
Tibshirani
R.
2010
Regularization paths for generalized linear models via coordinate descent
.
Journal of Statistical Software
33
(
1
),
1
.
Google Scholar
Crossref
Search ADS
PubMed
 
Gakii
C.
 &
Jepkoech
J.
2019
A Classification Model for Water Quality Analysis Using Decision Tree
.
Genuer
R.
,
Poggi
J. M.
,
Tuleau-Malot
C.
 &
Villa-Vialaneix
N.
2017
Random forests for big data
.
Big Data Research
9
,
28
–
46
.
Google Scholar
Crossref
Search ADS
 
Ghorbani
M. A.
,
Deo
R. C.
,
Yaseen
Z. M.
,
Kashani
M. H.
 &
Mohammadi
B.
2018
Pan evaporation prediction using a hybrid multilayer perceptron-firefly algorithm (MLP-FFA) model: case study in North Iran
.
Theoretical and Applied Climatology
133
(
3
),
1119
–
1131
.
Google Scholar
 
Giang
N. H.
,
Hieu
T. D.
 &
Do
H. N. T.
2021
Predictive Modelling Physico-Chemical Properties Groundwater in Coastal Plain Area of Vinhlinh and Giolinh Districts of Quangtri Province, Vietnam
.
Gutiérrez
Á. G.
,
Schnabel
S.
 &
Contador
J. F. L.
2009
Using and comparing two nonparametric methods (CART and MARS) to model the potential distribution of gullies
.
Ecological Modelling
220
(
24
),
3630
–
3637
.
Google Scholar
Crossref
Search ADS
 
Haghiabi
A. H.
,
Nasrolahi
A. H.
 &
Parsaie
A.
2018
Water quality prediction using machine learning methods
.
Water Quality Research Journal
53
(
1
),
3
–
13
.
Google Scholar
Crossref
Search ADS
 
He
Q.
,
Dong
Z.
,
Zhuang
F.
,
Shang
T.
 &
Shi
Z.
2012
Parallel decision tree with application to water quality data analysis
. In:
International Symposium on Neural Networks
(Jun Wang, Gary G. Yen, Marios M. Polycarpou, eds.).
Springer
,
Berlin, Heidelberg
, pp.
628
–
637
.
Google Scholar
Crossref
Search ADS
 
İlhan
N.
,
Yetiş
A. D.
,
Yeşilnacar
M. İ.
 &
Atasoy
A. D. S.
2021
Predictive modelling and seasonal analysis of water quality indicators: three different basins of Şanlıurfa, Turkey
.
Environment, Development and Sustainability
24(3),
1
–
35
.
Google Scholar
 
Jalal
D.
 &
Ezzedine
T.
2020
Decision tree and support vector machine for anomaly detection in water distribution networks
. In:
2020 International Wireless Communications and Mobile Computing (IWCMC)
(Azzam Mourad, Chadi Abou Rjeily, George Hadjichristofi, eds).
IEEE
, Beijing, China, pp.
1320
–
1323
.
Google Scholar
Crossref
Search ADS
 
Jaloree
S.
,
Rajput
A.
 &
Gour
S.
2014
Decision tree approach to build a model for water quality
.
Binary Journal of Data Mining & Networking
4
,
25
–
28
.
Google Scholar
 
Kanwal
S.
,
Gabriel
H.
,
Mahmood
K.
,
Ali
R.
,
Haidar
A.
 &
Tehseen
T.
2015
Lahore's groundwater depletion-A review of the aquifer susceptibility to degradation and its consequences
.
University of Engineering and Technology Taxila. Technical Journal
20
(
1
),
26
.
Google Scholar
 
Kardani
N.
,
Zhou
A.
,
Nazem
M.
 &
Shen
S. L.
2020
Estimation of bearing capacity of piles in cohesionless soil using optimised machine learning approaches
.
Geotechnical and Geological Engineering
38
(
2
),
2271
–
2291
.
Google Scholar
Crossref
Search ADS
 
Khaldi
R.
,
El Afia
A.
 &
Chiheb
R.
2019
Forecasting of BTC volatility: comparative study between parametric and nonparametric models
.
Progress in Artificial Intelligence
8
(
4
),
511
–
523
.
Google Scholar
Crossref
Search ADS
 
Khan
Y. A.
,
Shan
Q. S.
,
Liu
Q.
 &
Abbas
S. Z.
2020b
A nonparametric copula-based decision tree for two random variables using MIC as a classification index
.
Soft Computing
25(15),
1
–
16
.
Google Scholar
 
Khoshhal
J.
 &
Mokarram
M.
2012
Model for prediction of evapotranspiration using MLP neural network
.
International Journal of Environmental Sciences
3
(
3
),
1000
–
1009
.
Google Scholar
 
Kohler
M.
,
Krzyzak
A.
 &
Langer
S.
2019
Deep learning and MARS: a connection
.
Stat
1050
,
8
.
Google Scholar
 
Krutwagen
M.
2007
Impact of Shrimp Pond Wastewater on the Estuaries and the Issue of Salinity Intrusion in the Quang Tri Province
.
Bachelor's Thesis
,
University of Twente
.
Google Scholar
 
Kumar
C. P.
2012
Climate change and its impact on groundwater resources
.
International Journal of Engineering and Science
1
(
5
),
43
–
60
.
Google Scholar
 
LeBlanc
M.
 &
Tibshirani
R.
1994
Adaptive principal surfaces
.
Journal of the American Statistical Association
89
(
425
),
53
–
64
.
Google Scholar
Crossref
Search ADS
 
Loaiciga
H. A.
,
Charbeneau
R. J.
,
Everett
L. G.
,
Fogg
G. E.
,
Hobbs
B. F.
 &
Rouhani
S.
1992
Review of ground-water quality monitoring network design
.
Journal of Hydraulic Engineering
118
(
1
),
11
–
37
.
Google Scholar
Crossref
Search ADS
 
Loh
W. Y.
2011
Classification and regression trees
.
Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery
1
(
1
),
14
–
23
.
Google Scholar
 
Mukate
S.
,
Panaskar
D.
,
Wagh
V.
,
Muley
A.
,
Jangam
C.
 &
Pawar
R.
2018
Impact of anthropogenic inputs on water quality in Chincholi industrial area of Solapur, Maharashtra, India
.
Groundwater for Sustainable Development
7
,
359
–
371
.
Google Scholar
Crossref
Search ADS
 
Najafzadeh
M.
 &
Niazmardi
S.
2021
A novel Multiple-Kernel support vector regression algorithm for estimation of water quality parameters
.
Natural Resources Research
30(5),
1
–
15
.
Google Scholar
 
Najafzadeh
M.
,
Homaei
F.
 &
Farhadi
H.
2021a
Reliability assessment of water quality index based on guidelines of national sanitation foundation in natural streams: integration of remote sensing and data-driven models
.
Artificial Intelligence Review
54(6),
1
–
33
.
Google Scholar
 
Najafzadeh
M.
,
Homaei
F.
 &
Mohamadi
S.
2021b
Reliability evaluation of groundwater quality index using data-driven models
.
Environmental Science and Pollution Research
29(6),
1
–
17
.
Google Scholar
 
Najah
A.
,
El-Shafie
A.
,
Karim
O. A.
 &
El-Shafie
A. H.
2014
Performance of ANFIS versus MLP-NN dissolved oxygen prediction models in water quality monitoring
.
Environmental Science and Pollution Research
21
(
3
),
1658
–
1670
.
Google Scholar
Crossref
Search ADS
PubMed
 
National technical regulation on domestic water quality of Vietnam, QCVN 01: 2009/BYT
.
2021
.
(accessed 24 December 2021)
.
Omarova
A.
,
Tussupova
K.
,
Berndtsson
R.
,
Kalishev
M.
 &
Sharapatova
K.
2018
Protozoan parasites in drinking water: a system approach for improved water, sanitation and hygiene in developing countries
.
International Journal of Environmental Research and Public Health
15
(
3
),
495
.
Google Scholar
Crossref
Search ADS
 
Organisation mondiale de la santé, Światowa Organizacja Zdrowia, World Health Organization, & World Health Organisation Staff
2004
Guidelines for Drinking-Water Quality
, Vol.
1
.
World Health Organization
.
Pekel
E.
2020
Estimation of soil moisture using decision tree regression
.
Theoretical and Applied Climatology
139
(
3
),
1111
–
1119
.
Google Scholar
 
Ramchoun
H.
,
Idrissi
M. A. J.
,
Ghanou
Y.
 &
Ettaouil
M.
2016
Multilayer perceptron: architecture optimization and training
.
International Journal of Interactive Multimedia and Artificial Intelligence
4
(
1
),
26
–
30
.
Google Scholar
Crossref
Search ADS
 
Saghebian
S. M.
,
Sattari
M. T.
,
Mirabbasi
R.
 &
Pal
M.
2014
Ground water quality classification by decision tree method in Ardebil region, Iran
.
Arabian Journal of Geosciences
7
(
11
),
4767
–
4777
.
Google Scholar
Crossref
Search ADS
 
Sahu
S. K.
,
Dey
D. K.
 &
Branco
M. D.
2003
A new class of multivariate skew distributions with applications to Bayesian regression models
.
Canadian Journal of Statistics
31
(
2
),
129
–
150
.
Google Scholar
Crossref
Search ADS
 
Sekulic
S.
 &
Kowalski
B. R.
1992
MARS: a tutorial
.
Journal of Chemometrics
6
(
4
),
199
–
216
.
Google Scholar
Crossref
Search ADS
 
Sunardi, S., Ariyani, M., Withaningsih, S., Darma, A. P., Wikarta, K., Parikesit, P., Kamarudin, M. K. A. & Abdoellah, O. S.
2001
An alternative to neural nets: multivariate adaptive regression splines (MARS)
.
PC AI
15
(
1
),
38
–
41
.
Sunardi
S.
,
Ariyani
M.
,
Withaningsih
S.
,
Darma
A. P.
,
Wikarta
K.
,
Parikesit
P.
,
Kamarudin
M. K. A.
 &
Abdoellah
O. S.
2021
Peri-urbanization and sustainability of a groundwater resource
.
Environment, Development and Sustainability
23
(
6
),
8394
–
8404
.
Google Scholar
Crossref
Search ADS
 
Taylor
K. E.
2001
Summarizing multiple aspects of model performance in a single diagram
.
Journal of Geophysical Research: Atmospheres
106
(
D7
),
7183
–
7192
.
Google Scholar
Crossref
Search ADS
 
Touzani
S.
,
Granderson
J.
 &
Fernandes
S.
2018
Gradient boosting machine for modeling the energy consumption of commercial buildings
.
Energy and Buildings
158
,
1533
–
1543
.
Google Scholar
Crossref
Search ADS
 
Yang
J. H.
 &
Yang
M. S.
2005
A control chart pattern recognition system using a statistical correlation coefficient method
.
Computers & Industrial Engineering
48
(
2
),
205
–
221
.
Google Scholar
Crossref
Search ADS
 
Yurochkin
M.
,
Agarwal
M.
,
Ghosh
S.
,
Greenewald
K.
,
Hoang
N.
 &
Khazaeni
Y.
2019
Bayesian nonparametric federated learning of neural networks
. In
International Conference on Machine Learning
.
PMLR
, pp.
7252
–
7261
.
Google Scholar
 
Zheng
X.
,
Dan
C.
,
Aragam
B.
,
Ravikumar
P.
 &
Xing
E.
2020
Learning sparse nonparametric dags
. In
International Conference on Artificial Intelligence and Statistics
.
PMLR
, pp.
3414
–
3425
.
Google Scholar
 
© 2022 The Authors
This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/).