Application of Water Quality Indices, Machine Learning Approaches, and GIS to Identify Groundwater Quality for Irrigation Purposes: A Case Study of Sahara Aquifer, Doucen Plain, Algeria

Abstract

In order to evaluate and project the quality of groundwater utilized for irrigation in the Sahara aquifer in Algeria, this research employed irrigation water quality indices (IWQIs), artificial neural network (ANN) models, and Gradient Boosting Regression (GBR), alongside multivariate statistical analysis and a geographic information system (GIS), to assess and forecast the quality of groundwater used for irrigation in the Sahara aquifer in Algeria. Twenty-seven groundwater samples were examined using conventional analytical methods. The obtained physicochemical parameters for the collected groundwater samples showed that Ca²⁺ > Mg²⁺ > Na⁺ > K⁺, and Cl⁻ > SO₄²⁻ > HCO₃⁻ > NO₃⁻, owing to the predominance of limestone, sandstone, and clay minerals under the effects of human activity, ion dissolution, rock weathering, and exchange processes, which indicate a Ca-Cl water type. For evaluating the quality of irrigation water, the IWQIs values such as irrigation water quality index (IWQI), sodium adsorption ratio (SAR), Kelly index (KI), sodium percentage (Na%), permeability index (PI), and magnesium hazard (MH) showed mean values of 47.17, 1.88, 0.25, 19.96, 41.18, and 27.87, respectively. For instance, the IWQI values revealed that 33% of samples were severely restricted for irrigation, while 67% of samples varied from moderate to high restriction for irrigation, indicating that crops that are moderately to highly hypersensitive to salt should be watered in soft soils without any compressed layers. Two-machine learning models were applied, i.e., the ANN and GBR for IWQI, and the ANN model, which surpassed the GBR model. The findings showed that ANN-2F had the highest correlation between IWQI and exceptional features, making it the most accurate prediction model. For example, this model has two qualities that are critical for the IWQI prediction. The outputs’ R² values for the training and validation sets are 0.973 (RMSE = 2.492) and 0.958 (RMSE = 2.175), respectively. Finally, the application of physicochemical parameters and water quality indices supported by GIS methods, machine learning, and multivariate modeling is a useful and practical strategy for evaluating the quality and development of groundwater.

1. Introduction

Groundwater is a crucial source of freshwater for humans. However, with enormous population growth, agricultural use, industrialization, and unplanned urbanization, the extraction of groundwater has continuously increased from 283 km³/year in the 1960s to more than 700 km³/year in the last decades [1]. In many parts of the world, excessive groundwater extraction threatens both the amount and quality of water by causing dramatic drops in the water tables. Groundwater quality is essential in determining whether it is acceptable for use in agriculture and other industries. Numerous elements, like the makeup of the soil, seasonal variations, hydrogeochemical practices, and human actions, have an influence on the quality of groundwater [2,3,4,5,6].

In recent years, the deterioration of groundwater quality has been exacerbated by uncontrolled leaching of leachate from landfills, excessive use of fertilizers, and other human activities that are significantly dangerous to human health [7,8,9,10]. Many studies have evaluated groundwater quality in the world [11,12]. Nevertheless, the majority of these studies primarily concentrated on single-parameter evaluation, in which the governing elements were assessed individually, and the most compromised factor had a significant influence on water quality [13]. Hence, there is an increasing need for techniques that allow for better water quality interpretation to ensure effective groundwater quality control and management. Owing to their statistical ease of use and adaptability. Irrigation water quality indices (IWQIs) are among the best methods for monitoring groundwater quality for agricultural activities. IWQIs can help to interpret complex water quality data in simple terms [14,15,16]. Several indices have been used to categorize groundwater quality for agricultural purposes. These indices include the irrigation water quality index (IWQI), permeability index (PI), sodium adsorption ratio (SAR), Kelly index (KI), magnesium hazard (MH), and sodium percentage (Na%) [17].

Additionally, researchers’ interest in the application of multivariate statistical methods, such as principal component analysis (PCA) and cluster analysis (CA), as well as geographic information system (GIS) analysis, has grown [18], particularly in the assessment of water quality [19]. Multivariate statistical approaches and GIS analysis have been utilized collaboratively to highlight the primary variables affecting the geographical distribution of groundwater quality. For example, Nas and Berktay [20] used GIS and traditional kriging to analyze the spatial variation of groundwater quality indicators and map groundwater quality in urban areas in Konya, Turkey. Gilbert et al. [21] used a combination of GIS and multivariate analysis to analyze groundwater distribution in India. Studies conducted in multiple nations, including the United States, and India, revealed that groundwater found in various geological settings exhibits a variety of hydrochemical properties and falls into a few graded categories of water intended for irrigation [14,15,16,22,23,24,25,26].

In addition, by assigning weights to important ions based on entropy, researchers have explored approaches to reduce the subjectivity of existing water quality index technology, which has been displayed to be a more precise and vital method for precise weighing systems [27]. Water quality research, on the other hand, necessitates a significant amount of data collection, laboratory analysis, data management, and testing [28]. As a result of the computation’s subjectivity, WQI’s interpretation of the findings contains inconsistencies. According to previous studies, there is not a perfect WQI model. Consequently, it is crucial to implement a workable and affordable plan for accurate water quality assessment. Further, techniques for selecting features based on models [29] can be used to isolate a subset of features that have a predictive and a high discriminative. This method can improve model performance by reducing over-fitting and deleting unnecessary features. In addition to enhancing interpretability, retaining the original feature representation has its own set of advantages [30]. Moreover, there is a growing demand for feature selection algorithms in the fields of modeling and prediction [31]. Numerous researchers have investigated the effectiveness of various methods, such as Gradient Boosting Regression (GBR), Decision Tree (DT), and Random Forest (RF), for reducing the dimensionality of data. The RF model ranks all factors in terms of their importance [32]. Glorfeld [33] created a back-propagation neural network index to recognize the most important elements. In addition, the picking of hyperparameters has numerous benefits and can greatly affect the performance of any machine learning model. It could, for instance, boost ML algorithm performance [34] or make scientific studies more equitable and reproducible [35]. Given its potential significance in bettering the prediction model [36], it has a direct impact on how training algorithms act.

The groundwater of the Doucen plain is confined to the Sahara aquifers in the Northwestern region of Ouled Djelal, Algeria. It is exposed to varying degrees of salinity, which is attributed to the underlying fresh groundwater horizon, various pollutants of anthropogenic origin, and connate water trapped in aquifer rocks. However, to the best of the author’s knowledge, no comprehensive water quality research has been conducted on the Doucen plain. Physicochemical parameters focused on hydrochemical features offer preliminary knowledge regarding water facies, various geochemical pathways, and water classification [17,37]. Therefore, this study aims to (i) identify the groundwater characteristics, groundwater facies, and geochemical controlling mechanisms using physicochemical parameters, Chloro-alkaline indices (CAI 1 and CAI 2), multiple graphical approaches, and multivariate statistical analysis, (ii) assess and classify the groundwater quality for irrigation using IWQIs and (iii) assess the performance of ANN and GBR models to forecast the six IWQIs, namely IWQI, SAR, KI, Na%, MH, and PI.

2. Materials and Methods

2.1. Study Area

The region of Doucen is located in the Northwestern city (“wilaya”) of Ouled Djelal (Figure 1). The plain of Doucen extends over almost 12.42 miles from the wilaya of Ouled Djelal and 49.70 miles from the capital of the Biskra Provence. It is located at an elevation of over 102 m and between 4°57′ and 5°17′ east longitude and 34°30′ and 34°45′ north latitude. Area-wise, it’s 247.87 square miles, and it has administrative borders with the communes of Chaiba to the west, El-Ghrouss to the east and north, the town of Ouled Djalel to the south, the commune of Lioua to the east, and the commune of Still (El Oued willaya) to the south-east. It is an oasis with a strong agricultural vocation that mainly produces dates but also melons and watermelons. The commune of Doucen is vast, and it brings together several urbanized centers; the most important ones are “Doucen Ville” and “Tafechna”.

Geologically, the Doucen region consists of formations of ages ranging from Secondary to Quaternary. Secondary formations (Cretaceous) consist of limestone, crystalline limestone, dolomites, sandstone, gypsum, anhydrite, clays, and marls, while tertiary formations (Paleogene and Neogene) consist of limestone, marls, gypsum, clays, gravel, and sand red. The quaternary formations consist of scree, pebbles, gravel, sand, gypsum limestone, sandy alluvium, and clay alluvium (Figure 2). Hydrogeological studies have made it possible to highlight the existence of several aquifer reservoirs of distinct importance in terms of their lithological constitution, their geological structure, and the ease with which they are exploited (Figure S1).

2.2. Sampling and Hydrochemistry

2.2.1. Samples Collection

Twenty-seven (27) samples were collected from the study region (Figure 2). The positions of the sample stations were recorded using a global positioning system (GPS map, 76 CSx). During and after the wet season, contaminants are likely to be susceptible to seeping downhill [38]. Thus, water was pumped out for approximately 15 min before sampling to eliminate stagnant water. The samples were collected using polypropylene (PP) bottles, as proposed in a previous study [39]. Then, 0.45-μm acetate cellulose filters were employed to clean the samples. In the absence of acidification, 250 cm³ PE bottles were used for collecting samples for anions analysis. At last, the Algerian Waters laboratory in Biskra received the samples that had been kept at a temperature of <4 °C (The laboratory is accredited under ISO/IEC 17025).

2.2.2. Measurement of the Physicochemical Parameters

A WTW multiparameter (Weilheim, Germany) was used to monitor the physicochemical parameters (pH, temperature, and electric conductivity). Titration was employed to compute the concentrations of Ca²⁺, Mg²⁺, Cl⁻, and HCO₃⁻ ions. The value of SO₄²⁻ was found by employing spectrophotometry. In addition, a flame photometer was utilized in order to perform analyses on the K⁺ and Na⁺. TDS concentration was determined by weighing and drying at 103–105 °C in an oven. Nitrates were determined using the cadmium column reduction method. With an analytical uncertainty of less than 4%, all samples were examined in triplicate (Source: Gore, Michael. Spectrophotometry & Spectrofluorimetry. New York: Oxford University Press, 2000). The cations-anions balance was initially checked to make sure it was within 5%, which helped decide the analysis’ dependability by using recommended techniques from the American Public Health Association (APHA) [40]. The cadmium reduction method is a colorimetric method that involves contact of the nitrate in the sample with cadmium particles, which causes nitrates to be converted to nitrites. Hydrochemical outcomes were statistically investigated using DIAGRAMMES (version 5.8.0). STATISTICA software (version 8.0) was used to generate the results of the time series analysis and multivariate statistics.

2.2.3. Multivariate Statistical Analysis for Data Treatment

Large datasets can be streamlined and organized using multivariate statistical approaches to produce valuable insights [41,42]. In our case, two multivariate statistical methods, Viz. Cluster analysis (CA) and Principal component analysis (PCA) were applied to evaluate the physicochemical variables of our groundwater samples. Using a suite of multivariate techniques, cluster analysis can identify meaningful classes within a dataset. Unlike the non-hierarchical method, the hierarchical approach to CA does not require the entire number of clusters to be known in advance. According to a survey conducted by Belkhiri et al. [43], Ward’s method is superior to others because it results in a larger proportion of correctly identified observations. Thus, the present study adopts Ward’s clustering method, and the Euclidean distance was used in the current work to apply CA [44]. Minimal information loss was achieved by data reduction and the extraction of the most important components using PCA [44]. The aim of this project was to conduct the principal components analysis (PCA) method to extract the major elements related to the various causes of variance in hydrochemical data collected from the Doucen plain because PCA is used by Ridley. However, specifics were not shared [45].

2.3. Indexing Approaches

2.3.1. Chloro-Alkaline Indices (CAI 1 and CAI 2)

Multiple ionic connections were employed to assess the dynamic geochemical procedures in the current investigation. It comprises the dealings of [Ca²⁺ + Mg²⁺] vs. [SO₄²⁻ + HCO₃⁻], [Ca²⁺ + Mg²⁺] vs. [HCO₃⁻], [Ca²⁺/Na⁺] vs. [Mg²⁺/Na⁺], [HCO₃⁻/Na⁺] vs. [Ca²⁺/Mg²⁺] and [Na⁺] vs. [Cl⁻]. In addition, Equations (1) and (2) [46] were utilized to determine the ion exchange processes in the shallow aquifer using the Chloro-alkaline indices (CAI 1 and CAI 2):

$$\mathrm{CAI}1=\frac{{\mathrm{Cl}}^{-}-({\mathrm{Na}}^{+}+{\mathrm{K}}^{+)}}{{\mathrm{Cl}}^{-}}$$

(1)

$$\mathrm{CAI}2=\frac{{\mathrm{Cl}}^{-}-({\mathrm{Na}}^{+}+{\mathrm{K}}^{+)}}{{\mathrm{SO}}_4^{2-}+{\mathrm{HCO}}_3^{-}+{\mathrm{CO}}_3^{2-}+{\mathrm{NO}}_3^{-}}$$

(2)

2.3.2. Irrigation Water Quality Indices (IWQIs)

According to

Table 1. The IWQIs, related formula and reference.

WQIs	Formula	References
IWQI	${{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{Q}}_{\mathrm{i}}{\mathrm{W}}_{\mathrm{i}}$	[47]
SAR	$\frac{{\mathrm{Na}}^{+}}{\sqrt{\left({\mathrm{Ca}}^{2+}+{\mathrm{Mg}}^{2+}\right)/2}}$	[48]
KI	$\frac{{\mathrm{Na}}^{+}}{{\mathrm{Ca}}^{2+}+{\mathrm{Mg}}^{2+}}$	[49]
Na%	$\frac{\left({\mathrm{Na}}^{+}+{\mathrm{K}}^{+}\right)}{\left({\mathrm{Ca}}^{2+}+{\mathrm{Mg}}^{2+}\right)+\left({\mathrm{Na}}^{+}+{\mathrm{K}}^{+}\right)}\times 100$	[50]
MH	$\left[{\mathrm{Mg}}^{2+}/\left({\mathrm{Ca}}^{2+}+{\mathrm{Mg}}^{2+}\right)\right]\times 100$	[51]
PI	$\left(\frac{{\mathrm{Na}}^{+}+\sqrt{{\mathrm{HCO}}_3^{-}}}{\left({\mathrm{Ca}}^{2+}+{\mathrm{Mg}}^{2+}+{\mathrm{Na}}^{+}\right)}\right)\times 100$	[52]

Note(s): All IWQIs are calculated in meq/L.

, the IWQI, SAR, KI, Na%, MH, and PI% were all determined depended on the physicochemical parameters of the samples.

2.4. Spatial Distribution Pattern

GIS are important tools for the analysis and performance of spatial data associated with groundwater source control. Our topographic map was digitized using the USGS satellite and ArcGIS software (ver. 10.5.x) for the preparation of the base map of the study area. Garmin GPSMAP 64sx was used to pinpoint the precise locations of the sampling spots (i.e., bore hole locations) using the GIS platform. Using ArcGIS 10.5 software, we were able to produce maps presenting the geographical distribution of groundwater quality indicators, such as IWQI, SAR, KI, Na%, MH, and PI%, that are useful for irrigation.

2.5. Gradient Boosting Regression (GBR)

Decision trees are the building blocks of GBR and can be utilized for either regression or classification. A set of trees was constructed, with each tree focusing on the prediction residuals of the previous tree [53]. GBR’s hyperparameter tuning flexibility and ability to optimize a broad range of loss functions make it a highly flexible function-fitting technique. No pre-processing of the data is required, and it may be used for both numerical and categorical information. In addition, GBR can handle missing data on a large timescale. Simple trees are constructed in each iteration to prevent overfitting, and these can be used to extrapolate to novel data sets with greater precision. In general, a boosting approach has three parts: an additive model, some form of weak learners, and some type of loss function. The technique uses differentiable loss functions, can learn organically between input parameters over time, and can portray non-linear connections such as wind power curves [21]. To overcome the shortcomings of the underdeveloped models, gradient boosting machines are used to analyze gradients. To do this, we employ an iterative technique in which we unite base learners to decrease forecasting errors. Specifically, we use an additive model to combine decision trees and then employ gradient descent to decrease the loss function. The GBT (gradient boosting tree) F_n(x_t) can be stated as the summation of n regression-trees (Equations (3) and (4)).

$$F_n\left(x_t\right)={{\displaystyle \sum }}_{i=1}^n{f}_i\left(x_t\right)$$

(3)

where every f_i(x_t) is a decision tree (regression-tree). The ensemble of trees is constructed sequentially by estimating f_n+1(x_t) which is the new decision tree, using the following equation:

$$argmin{{\displaystyle \sum }}_{t}L(y_t.F_n\left(x_t\right)+f_{n+1}\left(x_t\right))$$

(4)

where L(.) is differentiable for loss-function L(.). A steepest descent method is used to solve it.

During GBR training, two criteria were considered: the number of boosting stages to perform (Ns) and the number of features to consider when looking for the best split (Mf). For Ns and Mf, the parameter values were (5,10, 15, 20, 25) and (‘auto’, ‘sqrt’, ‘log2’), respectively. Then, hyper-parameter optimization was performed, and the top-level model was built using the best values.

2.6. Back-Propagation Neural Network (BPNN)

The backpropagation neural network (BPNN) model is extremely popular [54]. There are three different types of layers that make up a BPNN: (1) the input layer, which serves as the neural net’s primary data source; (2) the concealed layer, which acts as a bridge between the independent inputs layer and the dependent output layer; and (3) the final layer, which outputs the results of the inputs that were given. Artificial neural networks (ANNs) are a layer-by-layer machine learning approach for extracting high-level features from unstructured data. The number of nodes in the network is proportional to how well the regression works, while the network itself contains a hidden layer. The hidden layers nodes contain a set of parameters, which are typically referred to as weight and biases, which are what constitute the “updated parameters” in the network. “Output layer” refers to the final layer, and it shows the estimated value of the target variables. By adopting a network of neurons or nodes connected by weighted connections, ANN models are instances of generalized mathematical modeling that replicate the human brain in pattern detection and prediction [55,56].

During training, the network was subjected to no fewer than 2000 iterations or until the error measurement was less than 10⁻⁴. In order to train the model and determine the optimal number of neurons (3, 6, 9, 12, 15, 18, and 21) for the model’s hidden layer, the validation strategy was combined with the LOOV technique. During the development of the ANN model, we investigated which activation function—”identity”, “logistic”, “tanh”, and “relu”—was the most effective. The restriction imposed by memory to efficiently implement the technique, the weight optimizer Broyden-Fletcher-Goldfarb-Shanno (lbfgs), was employed [57]. The most important feature has to be identified so that the regression model’s capacity for prediction may be increased and the complexity of hyperspectral images can be reduced [33,58].

2.7. Datasets and Software for Data Analysis

In this study, 27 samples were used across the training and validation processes. The model was trained and verified with the use of a technique known as leave-one-out cross-validation (LOOCV). In each experiment, LOOCV will keep most of the data for training purposes while setting aside a subset for validation. This approach may mitigate the effects of over-fitting and yield a more precise measure of the model’s prediction ability [59]. The software Python 3.7.3 was utilized for all these processes: data analysis, model construction, and data preparation. To perform tasks involving regression, the GBR and the ANN modules contained in the Scikit-learn package (version 0.20.2) were studied. A computer with an Intel Core i7-3630QM CPU running at a frequency of 2.4 GHz and 8 GB of Random Access Memory (RAM) was used for the data analysis.

2.8. Model Evaluation

The effectiveness of a regression model has been evaluated using the following statistical measures: the coefficient of determination (R²) and the root mean square error (RMSE) [60,61].

3. Results and Discussion

3.1. Groundwater Hydrochemical Properties

The analysis of the local elemental molar concentrations reveals that the cations evolved as follows: Ca²⁺ > Mg²⁺ > Na⁺ > K⁺ while the anions evolve in the following manner: Cl⁻ > SO₄²⁻ > HCO₃⁻ > NO₃⁻. The distribution of the main ions (Ca²⁺, Mg²⁺, and Cl⁻) is caused by anthropogenic factors such as irrigation water quality and unchecked fertilization; it is also strongly influenced by the local lithology. Both calcium and magnesium can come from the dissolution of calcium or magnesium sulphates. The transition from one dominant ratio to another can be done by dilution after mixing or precipitation of one of the ions. Regarding the origins of Cl⁻, it may result from the soil’s high levels of organic matter decomposing due to wastewater from “Doucen ville” [62].

Table 2. Descriptive outcomes of the groundwater samples with the standard limit for irrigation activities.

Parameters	FAO *	Minimum	Maximum	Average	Standard Deviation
pH	8.5	7.08	8.24	7.50	0.24
TDS	2000	1046	4650	2137	931
EC	3000	2091	9300	4274	1862
K+	2	2.00	50.00	16.89	13.20
Na+	919	50.00	340.00	167.11	89.98
Ca2+	400	176.00	561.00	353.48	110.30
Mg2+	60	75.00	423.00	155.44	84.31
SO42−	960	320.00	472.00	403.26	36.17
Cl−	1036	177.00	1808.00	621.67	384.37
HCO3−	610	450.00	830.00	548.81	79.74
NO3−	10	3.61	41.80	19.18	10.94

Note(s): FAO *: Food and Agricultural Organization [63].

displays a statistical summary of the groundwater samples that were analyzed in relation to irrigation water standards.

According to the physicochemical parameters results (Table 2), the pH values varied from 7.08 to 8.24 with a mean value of 7.50, which represents a weakly alkaline environment [64]. The pH values of the collected samples were within the allowable limit for irrigation. The TDS values diverse from 1045.5 mg/L to 4950 mg/L with an average value of 2136.81 mg/L (Table 2). Moreover, the electrical conductivity (EC) values ranged from 9300 µS/cm to 2091 µs/cm, with an average value of 4273.62 µS/cm (Table 2). According to the results of EC values, the majority of our samples (74%) were unfit for irrigation purposes. Table 2 shows that the average K⁺ content in our samples was 17.70 mg/L, with a range of 2–50 mg/L. This means that none of our groundwater samples are suitable for irrigation. Because of the high concentration of K⁺, we can deduce that potash feldspars minerals have been weathered and that chemical fertilizers (NPK) have been dissolved [65]. Further, Na⁺ concentrations extended from 50 to 340 mg/L, with an average of 168 mg/L, making all groundwater samples technically appropriate for irrigation (Table 2). The concentration of the Ca⁺² ion ranged from 176 to 561 mg/L, with a value of 353.48 mg/L serving as the average (Table 2). As shown by the Ca⁺² levels, the majority of the collected samples (78%) were acceptable for irrigation, while 22% were not. This discrepancy might be attributed to the impact of the dissolution of carbonate minerals from the gypsum deposits [66,67].

The concentration of Mg²⁺ ions in the collected water samples extended from 75 to 423 mg/L, with an average value of 155.44 mg/L (Table 2), signifying that all groundwater samples were inappropriate for irrigation. The elevated concentration of Mg²⁺ revealed leaching and ion exchange process of ferromagnesium minerals from the aquifer materials [68,69]. The HCO₃⁻ concentrations were from 450 to 830 mg/L, with an average of 548.81 mg/L. According to HCO₃⁻ values, the majority of samples (93%) were within the satisfactory limit for irrigation purposes consistent with the value guide of 610 mg/L [63]. According to [63], all of our groundwater samples were suitable for irrigation since their sulphate concentrations ranged from 320 to 472 mg/L, with an average of 403.29 mg/L (Table 2). Overall, the chloride concentrations in the groundwater samples that were taken diverse from 177 to 1808 mg/L, with an average of 621.66 mg/L (Table 2). Approximately 7% of the samples exceeded the required value related to irrigation, although the vast majority (93%) were under the permitted level for irrigation. Evaporitic formations, perhaps related to the local agro-food sector or urbanization via wastewater discharges [41], contributed to the elevated chloride concentrations found in the groundwater samples. Finally, the nitrate concentrations varied between 3.61 mg/L and 40.71 mg/L (average value of 20.48 mg/L), as indicated in Table 2.

More than two-thirds (67%) of the groundwater samples had nitrates values too high for irrigation, while 33% were completely unfit for the purpose.

The high concentration of NO₃⁻ suggested anthropogenic activity due to excessive inorganic nitrogenous fertilizer application, home and industrial wastes, and intense irrigation [52,64,65,70].

3.2. Groundwater Facies and Source Identification

3.2.1. Groundwater Types

Geology, water-rock interaction, and groundwater flow pattern through the aquifer are important in the production and classification of water types, which are, in turn, explained by groundwater hydrochemical facies. To better understand the preponderance of major ions in the groundwater aquifer, a Piper diagram was developed to depict the geochemical attribution and hydrogeochemical properties of groundwater [71]. Rock-water interaction, ion exchange, weathering of soils, and salt-bearing sedimentary rock were the main geochemical controlling processes, as shown in Figure 3.

3.2.2. Processes Influencing Groundwater Chemistry

The diagram of Gibbs can help in identifying the variables that affect the hydrochemistry of groundwater [72]. It classified the water into three types such as precipitation dominance, rock dominance, and evaporation dominance [73,74]. As shown in Figure 4, water-rock interactions appear to be the predominant natural process governing water chemistry, as evidenced by the fact that groundwater samples cluster in the rock weathering dominating field and the evaporation zone dominance. Furthermore, the ratios of Na⁺/(Na⁺ + Ca²⁺) were from 0.31 to 0.53; an average of 0.31 was recorded, which showed a robust cation exchange in the organization [47,75] according to Equations (5) and (6):

Ion exchange:

2NaX + Ca²⁺ → 2Na⁺ + CaX²

(5)

Reverse Ion exchange:

CaX² + 2Na⁺ → Ca²⁺ + 2NaX

(6)

The scatter plot HCO₃/Na versus Ca/Na revealed a high influence on the weathering process (Figure 5a). Generally, the dissolution of calcite minerals and dolomite was responsible for high Ca²⁺ and Mg²⁺ along with HCO₃⁻ ions in groundwater [76,77]. Moreover, the scatter plot between Mg/Na and Ca/Na demonstrated that the main mechanism and factor contributing to the presence of magnesium in groundwater is silicate weathering (Figure 5b). Nonetheless, the Ca²⁺ + Mg²⁺ vs. HCO₃⁻ + SO₄²⁻ relations of most groundwater samples (Figure 5c) revealed the non-dominancy of carbonate weathering in the study area. The plotting of Ca²⁺ + Mg²⁺ versus HCO₃⁻ (Figure 5d) indicated an excess of HCO₃⁻ due to rock weathering and the ion exchange process, which was a prime mechanism to release the HCO₃⁻ into the groundwater [78]. Consequently, Na⁺: Cl⁻ ratios in the majority of samples were lower than unity (Figure 5e). According to Figure 5e, the ratio of Na⁺/Cl⁻ was 0.28, indicating another chloride source [79]. Finally, chloro-alkaline indices (CAI) were applied to examine the cation exchange process and geochemical controlling mechanisms on the groundwater chemistry. The CAI values such as CAI 1 and CAI 2 showed positive values (Figure 5f), which indicates high cation exchange tendency between the Ca²⁺ and Mg²⁺ in the surrounding rock and Na⁺ and K⁺ in the groundwater [75,80,81].

3.3. Analysis of Multivariate Statistics

3.3.1. Cluster Analysis

Concerning the similarity of the groundwater samples, a combination of Ward’s linkage technique and Euclidean distance were used. In Figure 6, the Dendrogram was displayed to classify the different physicochemical variables in the obtained groundwater samples. In order to roughly match normally distributed data, all variables were log-transformed. Standard scores (z-scores) were calculated and applied to each variable [82].

The Dendrogram of the ten physicochemical parameters (Mg²⁺, Ca²⁺, K⁺, Na⁺, Cl⁻, HCO₃⁻, NO₃⁻, SO₄²⁻, EC and TDS) had been divided into three main groups (Figure 6). The obtained results revealed that TDS and EC were the major distinction parameters. The first group (G1) revealed a close association between the carbonate’s parameters HCO₃⁻, SO₄²⁻, Mg²⁺, and Ca²⁺, due to the major predominance of Mg²⁺, Ca²⁺ in the chemical composition of our groundwaters, such as sulphates or anhydride and calcium sulphates. The second group (G2) revealed a close association between the evaporate parameters. For example, Na⁺, K⁺, Cl⁻, EC, and TDS, indicate the main participation of chlorides and salts in the electrical conductivity with the dominance of Cl⁻ in the chemical composition of groundwater in the study region. Both G1 and G2 demonstrated that the lithological component dominated the mineralization of the Plio-Quaternary aquifer’s waters in Doucen. Finally, the third group (G3) showed a close association between Mg²⁺ and Ca²⁺, indicating the probable same origin of these two elements. G3 showed the dissociation of nitrates from other chemical elements present in groundwater, which indicates an anthropogenic source.

3.3.2. Principal Component Analysis (PCA)

Three factors (F) were retained from the PCA analysis with an Eigen value of more than one. F1 explained 59.86% of the dataset variability, while F2 and F3 explained, respectively, 11.35% and 10.35% (Figure 7). The value of the variable loading is presented in

Table 3. Correlation between the different physicochemical parameters and factors.

Parameter	F1	F2	F3
EC	0.965	0.051	−0.016
TDS	0.965	0.051	−0.016
Ca2+	0.781	−0.438	0.200
Mg2+	0.761	−0.448	0.121
HCO₃−	0.514	−0.007	0.551
Cl−	0.920	0.166	−0.148
Na+	0.724	0.522	−0.141
K+	0.805	0.444	−0.147
SO42−	0.761	−0.241	0.089
NO3−	−0.274	0.427	0.778

F1 (factor 1), F2 (factor 2), F3 (factor 3).

. The value close to ±1 revealed a high correlation between the factor and the variables. These loadings are further categorized into three categories, such as strong (>±0.75), moderate (±0.75 to ±0.50), and weak (±0.50 to ±0.30), as reported by Hinge et al. [44,74]. F1 has a robust positive relationship with TDS, EC, Cl⁻, Ca³⁺, Mg²⁺, K⁺, and SO₄²⁻. The source of SO₄²⁻ may be oxidation of SO₄²⁻ coming from fertilizer and sulfur compounds. The Ca²⁺, Mg²⁺, and K⁺ could be due to anthropogenic sources, such as irrigation water quality, domestic waste, and uncontrolled fertilization. The presence of chloride could be due to weathering of soils and salt-bearing formation. F2 was moderately correlated with Na⁺, while F3 was highly correlated with NO₃⁻ and moderate correlation with HCO₃⁻, which indicates alkaline water passing through rocks and soil. These results shed light on the methods through which human activities, rock weathering, and the ion exchange process affect groundwater quality.

3.4. Irrigation Water Quality Indices

Based on the parameters’ standard value intervals, we classified the quality of the irrigation water in our study area using the IWQI, SAR, KI, MH, Na%, and PI (

Table 4. Statistical analysis and classes of IWQIs.

Criteria	Min	Max	Mean	Range	Class	Number of Samples (%)
IWQI	19.01	73.61	47.17	85–100	No restriction	0 (0.0%)
				70–85	Low restriction	0 (0.0%)
				55–70	Moderate restriction	9 (33.33%)
				40–55	High restriction	9 (33.33%)
				0–40	Severe restriction	9 (33.33%)
SAR	0.63	4.10	1.88	<10	Excellent	27 (100%)
				10–18	Good	0.00
				19–26	Fair Poor	0.00
				>26	Unsuitable	0.00
KI	0.08	0.62	0.25	<1	Suitable	27 (100%)
				>1	Unsuitable	0.00
Na%	7.88	39.31	19.96	<20%	Excellent	16 (59.26%)
				21–40%	Good	11 (40.74%)
				41–60%	Permissible	0.00
				61–80%	Doubtful	0.00
				>80%	Unsuitable	0.00
MH	25.38	56.22	41.18	<50%	Suitable	25 (92.59%)
				>50%	Unsuitable	2 (7.41%)
PI	13.95	46.68	27.87	>75%	Suitable	0.00
				25–75%	Moderate	15 (55.55%)
				<25%	Unsuitable	12 (44.44%)

).

3.4.1. Irrigation Water Quality Index

IWQI involves the usage of either individual chemical indices [81,83,84,85] or several associated indices [47,86,87]. Although evaluating groundwater for irrigation based on individual characteristics is useful, the combined indices give decision-makers more insightful information.

In this study, five hazard parameters have been used: SAR, EC, Cl, Na, and HCO_3, to judge the groundwater’s viability for irrigation in agriculture [47]. The resulting IWQI value varies from 19.01 to 73.61, with an average value of 47.17. The IWQI suitability categorization of groundwater samples falls into three categories (moderate, high, and severe restriction) with an equal 33.33% (Table 4). Based on physicochemical parameters, the overall index map shows the water’s appropriateness for irrigation (Figure 8a). A map can be used to estimate groundwater validation for irrigation. According to IWQI values, the central region of our region has documented water quality degradation brought on by anthropogenic activities and geogenic sources (rock weathering and the ion exchange process).

3.4.2. Sodium Adsorption Ratio

The ability of the soil structure in the aquifer is shown by the sodium adsorption ratio. SAR is also employed in irrigation water to remove Ca²⁺ and Mg²⁺ ions and acquire Na⁺ ions from groundwater at ion-exchangeable sites, potentially causing soil particle dispersion and a decrease in infiltration capacity [88,89]. Whilst significant salinity of water can benefit the structure of the soil by improving the infiltration rate, which causes more water stress in plants. Table 4 shows that the calculated percent of SAR in groundwater samples varied from 0.63% to 4.1% (mean of 1.88%). According to sodium adsorption (SAR), four classes were created to identify the irrigation water [48], unsuitable (>26%), fair poor (19–26%), good (10–18%), and excellent (<10%). The value of SAR indicated that all of the study site’s water samples belong to the excellent category, which indicates its suitability for irrigation (Figure 8b). These results showed that irrigated water had no impact on crop output or soil infiltration.

3.4.3. Kelly Index

Groundwater suitability for irrigation was determined using the KI method, and the results indicated an overabundance of sodium ions in the water [90]. The KI value was diverse from 0.08 to 0.62, with a mean value of 0.25. With regards to the KI outcomes, 100% of the groundwater samples were deemed irrigation-ready (Table 4 and Figure 8c). If the KI value is greater than one (KI > 1), there is too much sodium present, whereas a number lower than one (KI < 1) denotes irrigation-ready water [91,92].

3.4.4. Sodium Percentage

Since Na⁺ concentrations have a negative effect on soil permeability, Na% can be utilized to define whether or not water is appropriate for irrigation purposes [93]. Clay minerals remove Mg²⁺ ions and Ca²⁺ from irrigation water by absorbing excess Na⁺. Na⁺ in water is exchanged for Ca²⁺ and Mg²⁺ in soil, reducing permittivity and, therefore, soil penetration. Sodium concentration ranged from 7.88% to 39.31%, with a mean of 19.95% across all samples (Table 4). There are five categories for the quality of irrigation water depending on the Na% value [94]: excellent (20%), acceptable (20–40%), allowed (40–60%), dubious (60–80%), and inappropriate (>80%). About 59.26% of the samples are in the excellent class, and these are positioned in the north, south east, and south west of the tested area. Nonetheless, about 40.74% of the samples are in the ideal category for irrigation and are located in the geographic center of the region (Figure 8d).

3.4.5. Magnesium Hazards

MH, defined as the ratio of magnesium to calcium in the water, was estimated using the above-mentioned equation (Table 1). As shown in the MH value, there are two main categories of groundwater utilization in irrigation. If the MH value is below 50%, the water is considered appropriate, whereas if it is above 50%, it is deemed unsuitable [95]. The rate at which soil absorbs water from irrigation may be affected by the adsorption of clay minerals to Mg²⁺ ions in water if the concentration of Mg ions is larger than that of calcium ions [89]. The groundwater samples showed the MH value varied from 25.38 to 56.22, with a mean value of 41.17. With an MH value of less than 50, groundwater was appropriate for irrigation in 25 samples, while with an MH value of more than 50, it was not (Figure 8e). These inappropriate samples are situated in the research area’s northern region.

3.4.6. Permeability Index

Since irrigation water consumption has a lasting effect on soil permeability, the PI serves as a vital gauge of irrigation water quality; it is regulated by bicarbonate, calcium, magnesium, and sodium ions found in the soil. To decide whether or not groundwater is suitable for irrigation [96], the parameter PI was used with the following cutoffs: unsuitable (PI = 25), good (PI = 25–75), and excellent (PI > 75). The PI value for our samples was reported to range from 13.95 to 46.68, with an average value of 27.86. Approximately 55.55% of the water samples were found in the moderate class, and the rest (44.44%) were unsuitable for irrigation; these samples were located in the north, south east, and south west of our area (Figure 8f).

4. IWQIs Prediction Using GBR and ANN

The high-level variables were filtered using the studied parameters, as shown in

Table 5. The outcomes of GBR and ANN models based on optimal features.

Variable	Model	Optimal Features (F)	Hyper-Parameters	Training		Validation
				R2	RMSE	R2	RMSE
IWQI	GBR	EC, Na+, Cl−	(Ns = 25, Mf = log2)	0.991	1.481	0.951	2.562
	ANN	Cl−, Na+	(h1 = 18, h2 = 9, fun = relu)	0.973	2.492	0.958	2.175
SAR	GBR	Mg2⁺, Na+	(Ns = 25, Mf = log2)	0.983	0.128	0.841	0.294
	ANN	Ca2⁺, Mg2⁺, Na+	(h1 = 9, h2 = 12, fun = logistic)	0.999	0.003	0.999	0.006
KI	GBR	Mg2⁺, Na+	(Ns = 25, Mf = log2)	0.981	0.020	0.763	0.054
	ANN	Mg2⁺, Na+, Ca2⁺	(h1 = 6, h2 = 15, fun = tanh)	0.999	0.002	0.999	0.004
Na%	GBR	Mg2⁺, Ca2⁺, Na+	(Ns = 25, Mf = auto)	0.990	0.909	0.783	3.186
	ANN	Ca2⁺, Na+	(h1 = 3, h2 = 18, fun = identity)	1.0	2.306 × 10−7	1.0	4.169 × 10−7
MH	GBR	Ca2⁺, Mg2⁺	(Ns = 25, Mf = auto)	0.969	1.125	0.557	2.946
	ANN	Ca2⁺, Mg2⁺	(h1 = 9, h2 = 9, fun = logistic)	0.999	0.005	0.999	0.036
PI	GBR	Na+, Mg2⁺	(Ns = 25, Mf = auto)	0.984	1.093	0.577	4.147
	ANN	HCO₃⁻, Mg2⁺, Ca2⁺, Na+	(h1 = 18, h2 = 15, fun = logistic)	0.999	0.003	0.999	0.106

Note(s): where Ns is the number of boosting stages to perform, Mf is the number of features to consider when looking for the best split, h₁ and h₂ are amount of neurons in the two hidden layers, and fun is the activation function.

and Figure 9, using the GBR and ANN models. These characteristics contributed to the identification of IWQI, SAR, KI, Na%, MH, and PI. The training and testing results (accuracy and RMSE) for forecasting the investigated parameters using the GBR and ANN models are publicized in Table 5 and Figure 9. The unused reserved values of GBR and ANN models were compared to the expected values. Results from the present study examination and comparison of multivariate methods show that this approach significantly improves predictability. Relying on independent validation to assess the accuracy of a regression model is preferable since validation data are not incorporated into the model construction practice. In this work, to forecast each output, the ANN model has surpassed the GBR model. The findings showed that the ANN-2F had the highest correlation between IWQI and the exceptional features, making it the most accurate prediction model. This model has roughly two qualities that are critical for IWQI prediction. The R² and RMSE values for the training and validation sets were 0.973, 2.492, and 0.958, 2.175, respectively. When measuring SAR, the ANN-3F model had the best outcomes. In the training set, the R² value was 0.999 (RMSE = 0.003), and in the testing phase, it was 0.999 (RMSE = 0.006). The most accurate model (ANN-3F) for determining KI had RMSE values for the training and testing sets of 0.002 and 0.004, respectively, and R² values of 0.999, and 0.999, respectively. Na% predictions were most accurate with the ANN-2F model. Regarding the training and test sets, the model achieved an R² of 1.0 (RMSE = 2.306 × 10⁻⁷) and 1.0 (RMSE = 4.169 × 10⁻⁷). The ANN-2F model accurately forecasted the MH. R² scores for the training and testing sets were 0.999 (RMSE = 0.005 and 0.036, respectively). In terms of predicting the PI, the ANN-4F model outscored other models. With R² values of 0.999 for the training set and 0.999 for the testing set (RMSE = 0.003 and 0.106, respectively), the model completed well. According to [97], who assert that the performance exceeded expectations, several techniques, including filtering high-level characteristics and changing model hyperparameters, were necessary to update regression approaches for correct prediction. In four main areas, deep learning algorithms exceed expectations: selecting the most relevant feature from an image’s color space, integrating image data with information about plants’ surroundings, augmenting data, and merging separate trained deep networks [98].

5. Conclusions

In this study, IWQIs, multivariate statistical analyses, GBR, and ANN models, backed by GIS techniques, were applied to determine groundwater quality for irrigation in the Doucen Plain, Northeastern Ouled Djelal, Algeria. According to the physicochemical results, groundwater in the study area revealed ion sequences of Ca²⁺ > Mg²⁺ > Na⁺ > K⁺, and Cl⁻ > SO₄²⁻ > HCO₃⁻ > NO₃⁻, which indicates Ca-Cl water type due to the dominance of sandstone, limestone and clay minerals under the influence of anthropogenic activities, rock weathering, ion dissolution, and exchange processes. The IWQI values revealed that 33% of our samples were severely restricted for irrigation, while 67% were diverse from moderate to high constraints for irrigation. In addition, SAR, KI, Na%, MH, and PI% showed that 100.0%, 100.0%, 59.0%, 92.0%, and 44.0% of groundwater samples were classified as excellent, suitable, excellent, suitable, and suitable, respectively. The ANN model has surpassed the GBR model. The findings showed that the ANN-2F had the highest correlation between IWQI and the exceptional features, making it the most accurate prediction model. For example, the most accurate model (ANN-3F) for determining KI had RMSE values for the training and testing sets of 0.002 and 0.004, respectively, with R² values of 0.999 and 0.999. Accordingly, an effective and practical strategy for assessing the quality of groundwater and its development is the use of physicochemical parameters and water quality indices supported by GIS techniques, multivariate modeling, and machine learning.