Abstract
The primary objective of this study was to calculate and assess the irrigation water quality index. Furthermore, an effective method for predicting IWQI using artificial neural network (ANN) and multiple linear regression (MLR) models was proposed. The accuracy performance of each model was evaluated at the end of this paper. According to the calculated index based on 49 groundwater samples, the Sidi El Hani aquifer was of good and sufficient quality. Moreover, both the ANN and MLR models performed well in terms of actual and predicted water quality. The ANN model, on the other hand, demonstrated the highest prediction accuracy. The results of this model also revealed that the predicted and computed values were close, with determination coefficients R2, RMSE, and MAE of about 0.95, 1.02, and 0.90, respectively. As a result, the proposed ANN model in this study was consistent and sufficient. These findings will help to guide irrigation water management decisions for the study aquifer in the future. The proposed ANN model can also be used to estimate the irrigation water index of other semi-arid aquifers, but accuracy is dependent on proper training techniques and selection parameters.
HIGHLIGHTS
Assessment of irrigation water quality (IWQI) index in a semi arid-environment.
Prediction of IWQI using ANN model.
Prediction of IWQI using MLR model.
Effectiveness of a machine learning tool (ANN) in accurately predicitng of IWQI.
Developing an accurate model may be valuable to manage the irrigation water quality.
Graphical Abstract
INTRODUCTION
Groundwater is an important component in many regions and locations around the world, and it plays an important role in the long-term development of various sectors. Unfortunately, groundwater quality has been severely degraded and threatened in recent decades as a result of a variety of factors such as rapid population growth, industrial and agricultural expansion, and climate change. These factors, in fact, reduce and destroy the value of aquifers as a source of water (Vaux 2011; Houéménou et al. 2020). Therefore, groundwater sustainability has become a major concern; the challenge is to protect these resources for future generations while also meeting the needs of many economic sectors, particularly agricultural systems (Elgettafi et al. 2013; Abdel-Fatteh et al. 2020).
Agriculture is the primary consumer of hydric resources worldwide, accounting for approximately 60% of available hydric resources, primarily in developed countries. Groundwater supplies, on the other hand, account for more than 90% of accessible resources in developing countries (Aliyu et al. 2017; Velasco-Muñoz et al. 2018). As a result, considering the quality of these resources for irrigation purposes is critical. The primary water quality concern in most irrigation situations is salinity content as well as soluble salt compositions (Zaman et al. 2018; Malakar et al. 2019; Mirzavand et al. 2020). A high salinity level can have a negative impact on crop yield, soil physical conditions, fertility requirements, and irrigation system performance (Adimalla 2018; Ramadan et al. 2019). Therefore, improving water quality is critical to ensuring the production of high-quality crops as well as the preservation of soil quality.
The sustainability of irrigation water quality is a growing field of study around the world. Indeed, various types of research have been carried out to thoroughly develop hydrochemical indices for evaluating irrigation water quality. Edmunds et al. (2003), for example, used geochemical and statistical approaches to define baseline concentrations to investigate groundwater's natural baseline quality. Li et al. (2013) used standard methods to assess groundwater quality in Pengyang county, China, including irrigation coefficient, sodium adsorption ratio (SAR), total alkalinity, and total dissolved solids (TDS). They also refined the approach for order preference by comparing it to the ideal solution model for full evaluation. Furthermore, many studies used a statistical technique to assess water quality. This method was crucial in assessing groundwater contamination and locating pollution sources. Venkatramanan et al. (2016) used statistical methods to identify the characteristics and variables influencing groundwater pollution in Miryang, Korea. The findings revealed that groundwater was impacted by salty water and nitrate pollution. Ettaib et al. (2017) analyzed the water quality of the Medjerda River using hydrographic techniques and the PHREEQC geochemical software, and its suitability for irrigation was determined based on its electrical conductivity (EC), SAR, and salt content. The study found that some points fell into the excellent to good and good to permissible irrigation water categories. The remaining ones, on the other hand, were labeled as questionable to unsuitable for irrigation, limiting river water use to plants. All of these techniques are incapable of producing generated data. As a result, significant efforts have been made to develop a water quality index (WQI). Karbassi et al. (2011), for example, created a WQI for the Gorganrood River based on a variety of river conditions and features, including geographical, hydrological, discharge rate, and pollution sources. Additionally, Singh et al. (2020) evaluated and studied the water quality of the Bharalu river using yearly datasets for its suitability for irrigation using the suggested entropy weighted for spatiotemporal variability. Furthermore, Parvez & Inayathulla (2020) developed a WQI for the upper Cauvery, Karnataka, India, by measuring five water quality parameters: pH, nitrates, chloride, EC, and fluoride.
IWQI, in particular, is a useful synthetic tool for arranging datasets. This index combines a number of physical and chemical variables, with a single number representing the level of water quality derived from a large number of water parameters (Mufeed et al. 2021). This method makes it very easy for planners to assess the quality and potential hazards of a water type based on a variety of factors (Muniz et al. 2020). Furthermore, IWQI allows for the evaluation and comparison of various water samples to avoid the negative impact on soil and plants (Abbasnia et al. 2018). The estimation of IWQI aids in lowering the cost of drilling wells for agricultural purposes in areas with high salinization of groundwater. As a result, it distinguishes between excellent groundwater areas for irrigation and those that are not (Maliki et al. 2020). Horton (1965) proposed the first modern IWQI, which sparked several types of research around the world. Among these studies, however, the calculation of IWQI varied from one researcher to the next. Meireles et al. (2010), for example, used multivariate statistical analysis to create a new IWQI tailored to the Brazilian context. This index considered five variables: EC, sodium, chloride, bicarbonate, and sodium adsorption ratio. Simsek & Gunduz (2007) also used a new GIS-integrated tool to evaluate irrigation water quality in relation to soil and crop problems. This index considered five hazard categories: salinity, infiltration and permeability, specific ion toxicity, trace element toxicity, and miscellaneous impact on sensitive crops. Furthermore, Misaghi et al. (2017) used a national sanitation foundation water quality index (NSFWQI) to evaluate irrigation water quality requirements (Ghezel Ozan River, Iran), which included seven chemical parameters: EC, sodium, chloride, pH, bicarbonate, sodium adsorption ratio, and TDS.
Recently, the intelligent approach has been widely used for assessing water quality in several research studies owing to its relevance to find a solution to a complex problem and to underline the relationship between input and output data. The intelligent approach is made up of several models, the most common of which are artificial neural networks (ANNs), which can adapt to any intermittent change. Abbaa et al. (2017), for example, used a combination of multilinear regression, ANN, and adaptive neuro-fuzzy interference system tools to forecast the dissolved oxygen concentration downstream of Agra city. Similarly, Khudair et al. (2018) assessed groundwater quality for drinking purposes in Baghdad using an ANN model. Furthermore, Alves et al. (2018) proposed a new alternative approach to determining the water quality index that combines ultra-visible spectrophotometry with an ANN. This combined approach may enable the prediction of WQI in areas where there is no infrastructure to assess WQI using traditional methods and to monitor water in real-time. Several studies have demonstrated the reliability of ANN in comparison to other tools; however, the ability of ANN approaches to predict IWQI has not been compared to conventional regression methods such as MLR.
In a similar context, the shallow Sidi El Hani aquifer, located in a semi-arid region of central-eastern Tunisia, is the primary source of water for irrigation. Nonetheless, over the past few decades, the measured salinity of this aquifer has risen from 6 g/l to 8 g/l, potentially leading to a reduction in agricultural production and, as a result, an economic downturn in the medium term. M'nassri et al. (2016) conducted a previous study in which they assessed the level of physical and chemical parameters in the investigated aquifer and mapped them using a GIS tool to evaluate their spatial distribution. In comparison to previous studies, none has presented data on IWQI and their prediction using ANN and MLR models. Hence, the purpose of this paper is to calculate the IWQI to assess the Sidi El Hani aquifer for irrigation purposes, using the methodology proposed by Meireles et al. (2010). In addition, we propose a method for predicting IWQI using ANN and multiple linear regression (MLR) models. The accuracy of predicted and computed IWQI values is then assessed using these two models. The findings will be useful as a biased technique for monitoring and predicting water quality change, allowing for better water resource management, planning, and decision-making regarding available resources, particularly in arid and semi-arid regions.
MATERIALS AND METHODS
Study area description
The Ouled Chamkeh plain is located in the Mahdia region, between the latitudes of 10°10′ and 10°25′ east and 10°12′ and 35°35′ north (Figure 1). This plain has a total area of 346 km2. The study area is bounded to the north and northeast by the sabkha of Sidi El Hani, to the east by the wadi Cherita, to the south by the sabkha of Cherita and the Ktitr Mountain, and to the west by the El Guessat Mountain. The research area has a mean annual precipitation of 270 mm and an evapotranspiration rate of 1,500 mm/year, making it a semi-arid region (INM 2016). Mio-Plio-Quaternary sediments cover the investigated area. These formations’ lithstratigraphic sequence is as follows: (i) Holocene deposits of clays, sands, silts, and evaporates; (ii) Pleistocene deposits of sand, halite hard limestone, silts, and soft white limestone with a thickness of about 120 m; and (iii) Mio-Pleistocene deposits of sand with an interlayered 50 m thickness. (Tagorti et al. 2013; M'nassri et al. 2019a). The shallow aquifer of Sidi El Hani is the main source of water in this area. The depth of this aquifer ranges from 3 m at the eastern of the area up to 26 m in the centre. The groundwater flow direction is from the west to the sabkha Sidi El Hani in the north-east and to the sabkha Cherita in the south (M'nassri et al. 2019b). The study area's economy is primarily based on agricultural activities, with more than 70% of the total area under cultivation. The main crops cultivated are olives and cereals.
Sample collection and laboratory analysis
During the field studies conducted in March 2015, 49 groundwater samples were collected (Figure 2). A multi-parameter portable quality meter was used to measure hydrogen ion activity (pH), temperature (T), and EC in situ. The samples were then kept at a temperature below 4 degrees Celsius until they were analyzed in the laboratory. The chemical analyses were carried out in accordance with the American Association of Public Health's standard methods (APHA 1995). Atomic absorption spectrometry was used to determine the concentrations of sodium and potassium. The titrimetric method was used to measure calcium and magnesium. The Mohr method was used to dose chloride concentrations. Titration with H2SO4 was used to determine bicarbonates. Colorimetric analysis was used to determine sulfate concentrations. TDS were determined by evaporating a pre-filtered sample until it was dry. The charge balance error for the samples was within the acceptable limit of 5% (Freeze & Cherry 1979), confirming the results’ accuracy. Furthermore, using the Equations (1)–(5) shown in Table 1, a set of fundamental parameters characterizing irrigation water quality, including SAR, percent Na, MAR, RSC, and PI, were calculated based on measured parameters. SAR and percent Na values that are too high reduce soil hydraulic conductivity and irrigation efficiency. If the SAR and percent Na are both greater than 10 meq/L and 80 percent, the water is deemed unfit for irrigation. The MAR ratio (Equation (3)) emphasizes the importance of magnesium for soil and plant health. It is considered an essential component when the MAR is less than 50. The RSC also emphasizes the abundance of carbonates and bicarbonates. These elements can have an effect on soil fertility when the RSC exceeds 1.5 meq/l. The PI was calculated using Equation (5). As a result, three classes were maintained. Good irrigation water suitability is defined as Classes I and II, with a maximum of 75%. Class III, on the other hand, includes points with a PI of less than 25% and water unfit for irrigation (M'nassri et al. 2018).
Calculated water quality parameter based on measured parameters
Quality parameter . | Formula adopted . | Reference . |
---|---|---|
Sodium adsorption ratio | ![]() | Richards (1954) |
Percentage of sodium | ![]() | Eaton (1950) |
Magnesium adsorption ratio | ![]() | Raghaunth (1989) |
Residual sodium carbonate | ![]() | Raghaunth (1989) |
permeability index | ![]() | Doneen (1964) |
Quality parameter . | Formula adopted . | Reference . |
---|---|---|
Sodium adsorption ratio | ![]() | Richards (1954) |
Percentage of sodium | ![]() | Eaton (1950) |
Magnesium adsorption ratio | ![]() | Raghaunth (1989) |
Residual sodium carbonate | ![]() | Raghaunth (1989) |
permeability index | ![]() | Doneen (1964) |
Computation of IWQI
The range and type of irrigation water for IWQI (Zahedi 2017)
IWQI range . | Water type . |
---|---|
85 − 100 | Excellent |
70–85 | Very good |
55–70 | Good |
40–55 | Satisfactory |
0–40 | Unsuitable |
IWQI range . | Water type . |
---|---|
85 − 100 | Excellent |
70–85 | Very good |
55–70 | Good |
40–55 | Satisfactory |
0–40 | Unsuitable |
ANN model
The ANN application created connections between inputs and outputs. The artificial neurons received weighted inputs through the neural structure, combined them, and then used the nonlinear operation to produce the output in the second step (Ehteshami et al. 2016; Foddis et al. 2019). The ANN is commonly referred to as a multi-layer perceptron (MLP), which can be either feedforward or feedback (Çelik et al. 2016). Different algorithms are used to train the MLP, the most popular of which is the back propagation algorithm (Qaderi & Babanejad 2017).
MLR model
The flow chart below depicts the methodology used to develop the IWQI, ANN, and MLR models (Figure 3).
Flow chart architecture of the methodology that was adopted to predict IWQI.
RESULTS AND DISCUSSION
Physical and chemical characteristics of irrigation water
Table 3 shows the minimum, maximum, mean, and standard deviation of TDS, EC, pH, major ions (Ca2+, Mg2+, Na+, K+, HCO3−, Cl−, and SO42−), and qualitative parameters like SAR, percent Na, MAR, RSC, and PI for all 49 samples. In this study, the ions concentration and calculated parameters are compared to the Food and Agriculture Organization's (FAO) criterion for agricultural purposes proposed by Ayers & Westcot (1994) and the classification proposed by the University of California (Doneen 1964). TDS levels in this study range from 2,400.0 to 8,410.0 mg/l, with a mean and standard deviation of around 4,280.0 and 1,493.0 mg/l, respectively. According to FAO guidelines, because all of the samples have TDS values greater than 2,000 mg/l, the water is subject to severe restrictions on its use in irrigation. Likewise, EC ranges from 3.75 to 12.00 mS/cm, with a mean of 6.10 mS/cm. All of the samples are classified as severe (>3 mS/cm). The use of this irrigation water may result in serious problems such as high soil salt levels, which can interfere with plant water absorption and have a serious negative impact on plant growth (Kumar & Shrivastava 2015). Furthermore, pH values range from 6.88 to 8.50, indicating neutral to slightly alkaline water, and all samples fall within the FAO's normal range of pH values (Ayres and Westcot).
Summary statistics of physical, chemical and qualitative parameters
Parameter . | Min. . | Max. . | Mean . | S.D* . | FAO usual range in irrigation water . |
---|---|---|---|---|---|
TDS (mg/l) | 2,400.0 | 8,410.00 | 4,280.00 | 1,493.00 | <2,000 |
EC (mS/cm) | 3.75 | 12.00 | 6.10 | 1.96 | <0–3 |
pH | 6.88 | 8.50 | 7.70 | 1.96 | – |
Ca2+ (meq/l) | 7.50 | 35.00 | 16.45 | 5.59 | <20 |
Mg2+ (meq/l) | 5.00 | 38.00 | 17.00 | 8.49 | <5 |
K+ (meq/l) | 0.36 | 1.80 | 0.66 | 0.28 | <0.05 |
HCO3− (meq/l) | 2.00 | 8.00 | 5.00 | 1.80 | <10 |
Na+ (meq/l) | 17.86 | 63.56 | 36.29 | 11.92 | <40 |
Cl− (meq/l) | 20.46 | 75.12 | 39.48 | 13.01 | <30 |
SO42− (meq/l) | 13.28 | 57.00 | 28.75 | 9.45 | <20 |
SAR | 5.03 | 16.41 | 8.24 | 2.89 | – |
%Na | 31.85 | 72.00 | 49.98 | 8.78 | – |
MAR | 25.00 | 73.53 | 47.67 | 12.35 | – |
RSC | −64.00 | −14.00 | −29.50 | 11.52 | – |
PI | 33.33 | 74.45 | 52.11 | 8.96 | – |
Parameter . | Min. . | Max. . | Mean . | S.D* . | FAO usual range in irrigation water . |
---|---|---|---|---|---|
TDS (mg/l) | 2,400.0 | 8,410.00 | 4,280.00 | 1,493.00 | <2,000 |
EC (mS/cm) | 3.75 | 12.00 | 6.10 | 1.96 | <0–3 |
pH | 6.88 | 8.50 | 7.70 | 1.96 | – |
Ca2+ (meq/l) | 7.50 | 35.00 | 16.45 | 5.59 | <20 |
Mg2+ (meq/l) | 5.00 | 38.00 | 17.00 | 8.49 | <5 |
K+ (meq/l) | 0.36 | 1.80 | 0.66 | 0.28 | <0.05 |
HCO3− (meq/l) | 2.00 | 8.00 | 5.00 | 1.80 | <10 |
Na+ (meq/l) | 17.86 | 63.56 | 36.29 | 11.92 | <40 |
Cl− (meq/l) | 20.46 | 75.12 | 39.48 | 13.01 | <30 |
SO42− (meq/l) | 13.28 | 57.00 | 28.75 | 9.45 | <20 |
SAR | 5.03 | 16.41 | 8.24 | 2.89 | – |
%Na | 31.85 | 72.00 | 49.98 | 8.78 | – |
MAR | 25.00 | 73.53 | 47.67 | 12.35 | – |
RSC | −64.00 | −14.00 | −29.50 | 11.52 | – |
PI | 33.33 | 74.45 | 52.11 | 8.96 | – |
The calcium content of the groundwater in the study area ranges from 7.50 to 35.00 meq/l, with a mean of 16.45 and a standard deviation of 5.59 meq/l. Furthermore, magnesium values range from 5.00 to 35.00 meq/l, with a mean value of 16.45 meq/l. These chemical parameters are primarily associated with the weathering of carbonate minerals such as calcite and dolomite, as well as chemical fertilizer leakage (M'nassri et al. 2019a; Aminiyan & Aminiyan 2020). According to FAO guidelines, the calcium levels in the groundwater samples are within an acceptable range (40 meq/l), making them suitable for irrigation (Ayers & Westcot 1994). The magnesium content in all samples, however, exceeds the FAO value limit of 5 meq/l. The potassium and bicarbonates concentrations range from 0.36 to 1.80 meq/l and from 2.00 to 8.00 meq/l, respectively. The maximum concentration of K+ is above the FAO recommended value (<0.05 meq/l). The K contents in groundwater have probably originated from agriculture fertilizers (Bekkoussa et al. 2013). Nevertheless, as per the FAO standard guideline for HCO3− (Ayers & Westcot 1994), all samples fall within the recommended permissible values for irrigation purposes (<10 meq/l).
As noted in Table 3, Na and Cl concentrations range from 17.86 to 63.56 meq/l and 20.46 to 75.12 meq/l, respectively. Previous studies (M'nassri et al. 2018, 2019a, 2019b) revealed that Cl and Na are the dominant ions in the groundwater of the studied area. They have several origins such as rock weathering and irrigation return flow. The usual levels of Cl and Na in irrigation water range from 0 to 30 and 0–40 meq/l, respectively (Ayers & Westcot 1994). In our study, nearly 80% of the samples exceed the permissible range for Cl; however, 37% of the samples have Na concentrations above 40 meq/l. It is acknowledged that the increase in Cl concentrations in irrigation water may affect crop growth and reduce productivity. As for Na contents, high levels lead to Na-soil exchange and, hence, generate soil dispersion as well as permeability reduction (Holgate et al. 2011). Table 3 also indicates that sulphate concentrations vary between 13.28 and 57.00 meq/l, with a mean value of 28.75 meq/l. Almost 85% of groundwater samples exceed the recommended range of sulphate for irrigation water (Ayers & Westcot 1994). High SO42− contents in irrigation water may generate soil saturation with gypsum and, thus, clog the pores. These changes may affect the physical characteristics of the soil such as permeability and aeration (Prapadopoulos 1984).
Table 3 also reports the qualitative parameters of the irrigation water. SAR computed values range from 5.03 to 16.41, with a mean and standard deviation of approximately 8.24 and 2.89, respectively. M'nassri et al. (2016) revealed that all the samples have a SAR value ranging between 0 and 18 categorized them as ‘excellent’ and ‘good’ classes for irrigation. As per % Na, the calculated values vary between 31.85 and 72.00. A higher percentage of Na (>60%) can lead to soil physical properties deterioration (Palanisamy et al. 2020). In turn, MAR values range from 25.00 to 73.53. The permissible limit of MAR for irrigation water is 50 (Doneen 1964). In our studied area, 59% of the samples have a MAR value above 50 and, thus they are unsuitable for irrigation. The RSC and PI computed values range from −64.00 to −14.00 and 33.33 to 74.45, respectively. In terms of their MAR values, all the samples are appropriate for irrigation.
IWQI assessment
In the current study, The FA/ACP was applied to 49 samples and 15 variables, including TDS, CE, pH, Ca2+, Mg2+, Na+, K+, HCO3−, Cl−, SO42−, SAR, %Na, MAR, RSC, and PI. The KMO adequacy test and the Bartlett sphericity test indicate that the value are greater than 0.5 and lower than 0.001, respectively. Hence, the performed factorial model is adequate for this study. Based on the eigenvalue, which should be greater than 1, four factors are selected (Table 4). The first factor, C1, expressed more than 35.58% of the total variance. The second factor, C2, has a variance of 27.57%. However, C3 and C4 factors have a variance of 11.15% and 9.63%, respectively. It should also be noted that the cumulative variance of the load factors is 83.94%.
Percentage of selected components
Selected component . | Without rotation . | Rotation of component . | ||||
---|---|---|---|---|---|---|
Eigenvalue . | % total of variance . | % of cumulative variance . | Eigenvalue . | % total of variance . | % of cumulative variance . | |
C1 | 5.51 | 36.73 | 36.73 | 5.33 | 35.58 | 35.58 |
C2 | 4.31 | 28.79 | 65.52 | 4.13 | 27.57 | 63.16 |
C3 | 1.64 | 10.94 | 76.47 | 1.67 | 11.15 | 74.31 |
C4 | 1.12 | 7.47 | 83.94 | 1.44 | 9.63 | 83.94 |
Selected component . | Without rotation . | Rotation of component . | ||||
---|---|---|---|---|---|---|
Eigenvalue . | % total of variance . | % of cumulative variance . | Eigenvalue . | % total of variance . | % of cumulative variance . | |
C1 | 5.51 | 36.73 | 36.73 | 5.33 | 35.58 | 35.58 |
C2 | 4.31 | 28.79 | 65.52 | 4.13 | 27.57 | 63.16 |
C3 | 1.64 | 10.94 | 76.47 | 1.67 | 11.15 | 74.31 |
C4 | 1.12 | 7.47 | 83.94 | 1.44 | 9.63 | 83.94 |
As indicated in Table 5, C1, which represents the important factor affecting water quality, is strongly correlated with Cl-, TDS, EC, SO42−, Mg2+, Na+, and MAR as follow: Cl− (0.87), TDS (0.86), CE (0.83), SO42− (0.81), Mg2+ (0.79), Na+ (0.75), and MAR (0.60). The C1 factor is introduced as the ‘salinity’ component. The second factor has a strong correlation with %Na, Na+, SAR, and PI. C2 is considered a ‘sodicity’ component. The factor C3 has a significant loading value with Ca2+ (0.82), but C4 has a correlation with pH (0.69) and HCO3− (0.82). All of these variables are unrelated to C1. The C3 and C4 factors are considered as ‘alkalinity’ components.
Component loads for the physicochemical parameters
Parameters . | Component loads matrix . | |||
---|---|---|---|---|
C1 . | C2 . | C3 . | C4 . | |
TDS (g/l) | 0.86 | 0.14 | 0.04 | 0.20 |
CE (mS/cm) | 0.83 | 0.23 | 0.14 | 0.19 |
pH | 0.33 | 0.02 | −0.06 | 0.69 |
Ca2+ (meq/l) | 0.33 | −0.39 | 0.82 | 0.03 |
Mg2+ (meq/l) | 0.79 | −0.51 | −0.27 | 0.06 |
Na+ (meq/l) | 0.75 | 0.89 | 0.12 | 0.16 |
K+ (meq/l) | 0.22 | 0.04 | 0.32 | 0.15 |
HCO3− (meq/l) | −0.08 | 0.22 | 0.20 | 0.82 |
Cl− (meq/l) | 0.87 | 0.14 | −0.01 | 0.19 |
SO42− (meq/l) | 0.81 | −0.10 | 0.27 | −0.23 |
SAR | 0.37 | 0.89 | 0.07 | 0.13 |
%Na | 0.02 | 0.98 | −0.02 | 0.08 |
MAR | 0.60 | −0.31 | −0.77 | 0.04 |
RSC | −0.76 | 0.45 | −0.17 | 0.06 |
PI | −0.05 | 0.78 | −0.03 | 0.12 |
Parameters . | Component loads matrix . | |||
---|---|---|---|---|
C1 . | C2 . | C3 . | C4 . | |
TDS (g/l) | 0.86 | 0.14 | 0.04 | 0.20 |
CE (mS/cm) | 0.83 | 0.23 | 0.14 | 0.19 |
pH | 0.33 | 0.02 | −0.06 | 0.69 |
Ca2+ (meq/l) | 0.33 | −0.39 | 0.82 | 0.03 |
Mg2+ (meq/l) | 0.79 | −0.51 | −0.27 | 0.06 |
Na+ (meq/l) | 0.75 | 0.89 | 0.12 | 0.16 |
K+ (meq/l) | 0.22 | 0.04 | 0.32 | 0.15 |
HCO3− (meq/l) | −0.08 | 0.22 | 0.20 | 0.82 |
Cl− (meq/l) | 0.87 | 0.14 | −0.01 | 0.19 |
SO42− (meq/l) | 0.81 | −0.10 | 0.27 | −0.23 |
SAR | 0.37 | 0.89 | 0.07 | 0.13 |
%Na | 0.02 | 0.98 | −0.02 | 0.08 |
MAR | 0.60 | −0.31 | −0.77 | 0.04 |
RSC | −0.76 | 0.45 | −0.17 | 0.06 |
PI | −0.05 | 0.78 | −0.03 | 0.12 |
In the current research, the computation of the IWQI is based on the parameters that have a good load in the first component such as Cl−, TDS, EC, SO42−, Mg2+, Na+, and MAR. C1 is considered as the most significant component that explains the global variability in irrigation water quality. The limiting values of the physicochemical parameters considered in the calculation of IWQI are determined on the basis of the FAO's recommended usual range in irrigation water (Ayers & Westcot 1994) and of the classification proposed by the University of California (Doneen 1964). Furthermore, the weight values are estimated based on the variance of the first component related to its explainability towards each parameter. The normalized weight values are presented in Table 6.
Calculated relative weight of each parameter
Parameter . | Weight . |
---|---|
Cl | 0.158 |
TDS | 0.156 |
CE | 0.151 |
SO42− | 0.146 |
Mg2+ | 0.142 |
Na+ | 0.135 |
MAR | 0.112 |
Total | 1.000 |
Parameter . | Weight . |
---|---|
Cl | 0.158 |
TDS | 0.156 |
CE | 0.151 |
SO42− | 0.146 |
Mg2+ | 0.142 |
Na+ | 0.135 |
MAR | 0.112 |
Total | 1.000 |
The IWQI computed values range from 19.29 to 55.41. About 40% of samples fell to a good category, and the rest of the samples are satisfactory. The last class can be used for coarse-textured soils that are characterized by high permeability (Abbasnia et al. 2018). In addition, the water can be applied only for plants with high salinity tolerance and with special salinity control practice, except for water with low Na, Cl, and HCO3 values (Yildiz & Karakus 2020).
ANN and MLR analysis and modeling
In the development of the ANN model, several different ANN architectures, weights, and parameters are explored until the error between modelled and measured output values dataset was minimized. In this study, a model with three neurons in the hidden layer shows the best results. Consequently, the network architecture selected is composed of seven input layers, a hidden layer that contains three neurons, and an output layer with one neuron. The final trained weights and biases of a random train with the optimum parameters are illustrated in Table 7.
Final trained weights and biases in the control run
. | Weights and biases in the input-hidden layers . | ||
---|---|---|---|
. | N1-1 . | N1-2 . | N1-3 . |
Cl | −1.45 | 0.27 | −0.21 |
TDS | −0.62 | −0.06 | −0.01 |
EC | −0.54 | 0.78 | −1.43 |
SO4 | −0.78 | −1.15 | −0.21 |
Mg | −0.57 | 0.14 | −0.70 |
Na | −1.96 | −1.12 | 0.76 |
MAR | −0.61 | −0.11 | 0.46 |
Bias | 1.09 | 0.27 | 0.17 |
. | Weights and biases in the hidden-output layer . | ||
IWQI | −2.15 | −0.94 | 2.66 |
Bias | − 0.39 |
. | Weights and biases in the input-hidden layers . | ||
---|---|---|---|
. | N1-1 . | N1-2 . | N1-3 . |
Cl | −1.45 | 0.27 | −0.21 |
TDS | −0.62 | −0.06 | −0.01 |
EC | −0.54 | 0.78 | −1.43 |
SO4 | −0.78 | −1.15 | −0.21 |
Mg | −0.57 | 0.14 | −0.70 |
Na | −1.96 | −1.12 | 0.76 |
MAR | −0.61 | −0.11 | 0.46 |
Bias | 1.09 | 0.27 | 0.17 |
. | Weights and biases in the hidden-output layer . | ||
IWQI | −2.15 | −0.94 | 2.66 |
Bias | − 0.39 |
Criteria for assessing the predictive validity of the ANN and MLR models
Statistical criteria . | ANN . | MLR . |
---|---|---|
R2 | 0.92 | 0.81 |
RMSE | 1.02 | 1.20 |
MAE | 0.90 | 1.63 |
Statistical criteria . | ANN . | MLR . |
---|---|---|
R2 | 0.92 | 0.81 |
RMSE | 1.02 | 1.20 |
MAE | 0.90 | 1.63 |
Additionally, the relationships between observed values and computed values generated from the ANN and MLR models, shown in Figure 4(a) and 4(b), indicate that several points lie on or close to the straight line. The R2 value is found to be 0.928 in the ANN model prediction; however, it is in the range of 0.818 in the MLR model prediction.
Correlation between predicted and observed IWQI values with (a) the ANN model; (b) the MLR model.
Correlation between predicted and observed IWQI values with (a) the ANN model; (b) the MLR model.
On the other hand, the comparison between the IWQI observed values and the ANN and MLR prediction results for all samples is revealed in Figure 5. It is clear that the ANN predicted values of IWQI are nearer to the observed values compared to the MLR predicted values, thus suggesting the consistency and adequacy of the proposed ANN model. Based on these findings, the ANN modeling approach can be applied as a predictive tool to point up the most appropriate values of IQWI under various anthropogenic and natural factors affecting water quality and, hence, it is promising for modeling monthly water quality efficiency by integrating the time series analysis.
Based on these findings, the use of groundwater of the Sidi El Hani basin, which is mostly in the poor water class, for irrigation purposes is very hazardous for soil and plants. Indeed, the most significant salinization source in this region is agricultural activities. Therefore, potential measures should be implemented to avoid soluble salts leaching that threaten water quality. This situation should be performed by controlling the anthropogenic and natural factors that affect the water quality, such as soil characteristics, land use type, fertilizers, livestock number, and groundwater level. At this point, the developed ANN model promise for modeling monthly water quality efficiency by integrating the time series analysis of the factors distressing the irrigation water quality. This suggestion allows, thus, for better monitoring of salt leaching, and sustainable management of groundwater.
CONCLUSION
In arid and semi-arid regions, groundwater is commonly the only source of irrigation. Therefore, knowledge and assessment of the irrigation water quality index are helpful tools to manage water resources. The current study aims to determine the IWQI of Sidi El Hani aquifer, located in central-eastern Tunisia, and to predict this index, using the ANN and MLR models to highlight the most adequate predicted values. The sufficient model can be generalized to other regional territories that are dominated by an arid or semi-arid climate and thus contribute to the sustainable use and management of water resources.
Furthermore, the ANN modeling results show that the model has a good determination coefficient (R2) that is equal to 0.92. The RMSE and MAE are also found to be 1.02 and 0.90, respectively. Additionally, the MLR model provides a satisfactory IWQI forecast. The statistical indicators such as R2, RMSE, and MAE are about 0.81, 1.20, and 1.63, respectively. Nevertheless, the comparison between computed IWQI and predicted values, with the ANN and MLR models, shows that the ANN model seems to prove its prediction sufficiency. Thus, based on these results, more exhaustive monitoring and implemented management practice of groundwater in the Sidi El Hani aquifer are compulsory.
ACKNOWLEDGEMENTS
This work is part of a research project entitled ‘Salinization of groundwater of the Sidi El Hani aquifer.’ It was funded by the Institution of Research and Higher Agricultural Education. The authors sincerely thank the CRDA of Mahdia for their help throughout this work.
ADDITIONAL INFORMATION
The authors declare no competing interests.
DATA AVAILABILITY STATEMENT
All relevant data are included in the paper or its Supplementary Information.