Groundwater quality assessment for agricultural utilizing indexical and machine learning techniques in Ouled Djellal Aquifer, Southern Algeria

Abstract

Groundwater represents the main water resource for irrigation in the Ouled Djellal region (southeast of Algeria). Despite the importance of groundwater in this area, its quality and sustainability remain insufficiently studied. Therefore, this study aimed to introduce an integrated analytical framework by combining multivariate statistical techniques i.e., Principal Component Analysis (PCA) and Hierarchical Ascending Classification (HAC), irrigation indices (IWQI, SAR, Na%, SSP, PS, and RSC), and machine learning (ML) models such as Artificial Neural Network (ANN), Support Vector Machine (SVM), and Multiple Linear Regression (MLR) to assess and predict groundwater quality for irrigation. The main difference with previous studies is the fact that this work applied Empirical Bayesian Kriging Regression Prediction (EBKRP) to spatialize irrigation indices derived from ML with higher precision. The approach enables cross-validation of model performance and captures complex nonlinear interactions among hydrochemical parameters. The attained results revealed that groundwater quality was varied from moderate to poor for irrigation, driven mainly by salinity and sodicity effects. In addition, the ANN model achieved the highest predictive accuracy (R² = 0.97, RMSE = 1.50), confirming its superiority in modelling complex hydrochemical behavior. The proposed modelling framework represents a methodological advancement for data-scarce arid regions, serving as a practical tool adaptable to groundwater monitoring and irrigation planning in similar regions.

Introduction

In countries with limited surface water resources, the richness of groundwater becomes determinant. It constitutes a vital resource for domestic supply, agriculture, and other human activities, particularly in arid and semi-arid regions^1,2. As a result, access to higher-quality water remains a fundamental priority^3,4. Consequently, understanding the geochemistry of groundwater is essential for elucidating the mechanisms underlying water chemistry acquisition, in order to counteract any actions that may qualitatively degrade this valuable resource^5,6. Nevertheless, in arid (and semi-arid) zones, water supply depends largely on groundwater aquifers, which are often threatened by the deterioration of their quality because of both anthropogenic and natural factors^7,8. Thus, assessing the geochemical profile of groundwater resources is a necessary step towards their responsible and sustainable management^9,10. The chemical composition of water reflects a multitude of complex interactions, including water-rock interactions within the geological matrix, residence time, and flow dynamics^11,12,13. These processes represent key mechanisms in the acquisition of groundwater chemistry, and a better knowledge of them will facilitate more efficient exploitation and adequate protection of the resource^14,15.

The development of multivariate statistical methods provides powerful tools for identifying the principal factors governing groundwater chemistry and contamination, thereby offering valuable support for the water resources sustainable management^16,17. In this work, hierarchical cluster analysis (HCA) and principal component analysis (PCA) are applied to groundwater quality from the Ouled Djellal region in southeastern Algeria, to assess water quality and delineate areas potentially exposed to pollution. The resulting findings, using Geographic Information Systems (GIS), enable the mapping of contamination extent and a better understanding of the spatial distribution of key water quality parameters^18,19. Besides, water quality index (WQI) has also emerged as a widely used indicator, distilling complex hydrochemical data into a single, synthetic score that simplifies interpretation for the public and decision-makers. This index is particularly useful tool to assess the suitability of groundwater for irrigation purposes²⁰. Consequently, machine learning (ML) is becoming increasingly important in the assessment and prediction of water quality (WQ) by offering faster, more accurate, economical, and advantageous predictions compared to traditional methods, even though the latter often rely on complex, incomplete, and sometimes limited data to a few parameters^21,22. This is specifically beneficial in situations where analytical resources are generally limited. For example, recent research showed that the use of ML models such as artificial neural networks (ANN) and multivariate regression (MR) allows for an effective estimation of the WQ^23,24. Because ML algorithms handle large, diverse datasets (including sensor measurements and satellite imagery), which enable rapid detection of water-quality changes. This facilitates quick identification of pollution sources (industrial or agricultural)²⁵, and supports more proactive management of water resource²⁶. Techniques like ANN, MLR, SVM, and others, have recently demonstrated their high performance in predicting WQIs, achieving determination coefficients exceeding 0.97 and extremely high classification accuracy²⁷. The application of ML also allows for the implementation of real-time alert and continuous evaluation systems for WQ, which facilitates water resources integrated management²⁸. Recent publications in the field of ML, indicated that the use of this technology has now become essential for modernizing water quality forecasting²⁹.

In recent years, the Ouled Djellal region has experienced increasing water scarcity³⁰. Boreholes are running dry, and irrigated water is lacking, and this situation threatens local agriculture particularly date palm cultivation making groundwater reserves a vital pillar for local communities, livestock, and vegetation. Nevertheless, the current state of groundwater, both in terms of quality and quantity, jeopardizes the sustainability of these populations and their primary sources of livelihood. Table 1 summarized an important recent works on water quality assessment and irrigation suitability which highlighted their methodology (model used), main findings, and limitations.

Table 1 Summary of recent works on water quality and irrigation suitability (methodology, findings, and limitations).

In order to evaluate irrigation water suitability in the Ouled Djellal region, this study introduce the integration between water quality indices (WQIs) and GIS-based machine learning approaches. Therefore, this study was conducted to (i) evaluate the influence of natural processes and anthropogenic pressures on groundwater quality in Ouled Djellal using physicochemical parameters and groundwater facies, Chloro-alkaline indices (CAI 1 and CAI 2), multiple graphical and statistical approaches, (ii) assess irrigation water quality using IWQIs, and (iii) develop and test machine learning models (ANN, SVM, MLR) integrated with GIS to predict and interpret irrigation suitability. The results of this study can guide local water management authorities in adopting sustainable irrigation practices, while reducing potential risks to human health and the surrounding ecosystems.

Materials & methods

Research area

The town of Ouled Djellal is located in the southeastern part of Algeria, bounded to northeast by the wilaya of Biskra. It covers an area of ​​ about 326.6 km². It is border by the municipality of Doucen and the municipality of Chaiba in the north and northwest and by the commune of besbes and the commune of Sidi Khaled in the south and south-west. While it is border by the wilaya of El-Oued in the east (Fig. 1). From a topographic point of view, Ouled Djellal is part of the pre-Saharan region. Its relief is not very rugged and is characterized by the dominance of large areas, the average altitude is about 200 m. Hydrographically, the town is drained to the south by an important temporary wadi (Wadi Djedi). The wadi’s flow is irregular and practically zero. Other less important wadis also pass through the area of Oued Besbes, Oued Diel and Oued Rtem. Regarding the climate, the Ouled-Djellal region belongs to the Mediterranean climate type with a Saharan bioclimatic influence. It is characterized by hot and dry summers (temperature between 35 °C and 45 °C during the day, and between 25 °C and 35 °C at night), and cold, dry winters (temperature between 10 and 20 °C during the day, and between − 2 and 5 °C at night). Rainfall is very low and irregularly distributed in time and space, with the average rainfall (annual) approximately 126.2 mm, which reflecting arid to semi-arid conditions.

Fig. 1

Map of the study area (Ouled Djelall, North Sahara Algeria). Map created using ArcGIS Pro 2.8.8 (Esri; https://www.esri.com/arcgis/about-arcgis).

Geological and hydrogeological setting

Geological and hydrogeological studied have made³⁴ to highlight the different types of geological formations. Several aquifer reservoirs exist, with a significant variation in importance because of the differences in their geological structure, lithological composition, and the ease of their exploitation (Fig. 2). The Ouled Djellal region belongs to the vast sedimentary basin of the Lower Sahara, specifically in the Zab plain, which located at the junction between the southernmost tip of the Saharian Atlas, the northern Saharian basin, and the southwestern edge of the Aurès. Structurally, the region juxtaposes two sectors as the following: (i) from the north, a folded zone of the Saharian Atlas with SW-NE oriented structures inherited from the Alpine orogeny; (ii) from the south, the Saharian collapse zone characterized by depressions filled with recent formations. The transition between these domains is marked by a large tectonic structure activated after the uplift of the Atlas, called the Saharian flexure or southern Atlas flexure. Ouled Djellal is predominantly composed of sedimentary formations, extending from the Barremian stage of the Cretaceous period at the base up to Quaternary deposits at the surface. The Lower Sahara actually represents a broad infill plain that has gradually subsided from the Lower Cretaceous through to the Quaternary. Indeed, this basin is primarily filled with post-Eocene continental Tertiary sediments, mainly conglomeratic sands interbedded with clay layers and clayey-sandy beds. Finally, the Neogene formations, originating mainly from the erosion of the Atlas Mountains, entirely conceal the underlying folded geological structures^35,36,37, as show in Fig. 3.

The lithostratigraphic column of the Ouled Djellal region includes, from the oldest to the most recent: alterations of gray or white limestones and marls, sometimes gypsiferous, from the Lower and Upper Senonian. Next are lagoonal deposits and dolomitic, locally saline passages from the Middle Eocene (Lutetian), sandy and clayey levels with pebbles characterizing the terminal complex of the Miocene-Pliocene. Finally, there are accumulations of gravel, sand, and clay; filling of basins and wadi corridors from the Quaternary. The Quaternary terrains, widely represented, are the foundations of the current soils and the main local aquifers, particularly in the sectors of Wadi Djedi^36,38, as seen in Fig. 2.

The Ouled Djellal region is underlain by a shallow Quaternary aquifer hosted in unconsolidated alluvial deposits that fill the valley of Wadi Djedi (Fig. 3). Groundwater recharge occurs primarily through infiltration from ephemeral flood flows and infiltration along the wadi beds after rainfall events. The aquifer, characterized by intergranular porosity, presents heterogeneous lithological facies composed mainly of sands, gravels, and silty layers, which locally control its hydraulic behavior. Its thickness is variable, reaching several tens of meters within the central valley. The piezometric surface follows the morphological axis of Wadi Djedi, indicating a flow direction from the piedmont towards the downstream plain with a gentle hydraulic gradient of about 4.3 × 10⁻². Pumping tests indicate a transmissivity on the order of 1.2 × 10⁻³ m²/s, consistent with moderate yield capacities from shallow boreholes ranging between 1.2 and 10 L/s. Although this aquifer is not highly productive, it constitutes the principal groundwater reservoir in the region, meeting domestic and agricultural demands. Nevertheless, the Quaternary aquifer remains sensitive to water-level declines due to limited natural recharge in this arid environment and the high extraction pressure during irrigation periods.

Fig. 2

Geological map of Biskra containing our water samples. Map created using ArcGIS Pro 2.8.8 (Esri; https://www.esri.com/arcgis/about-arcgis).

Fig. 3

Schematic geological cross-section of the study area (made by the authors). Map created using ArcGIS Pro 2.8.8 (Esri; https://www.esri.com/arcgis/about-arcgis).

Analysis and sampling of water

Twenty-three Groundwater (GW) samples from the study region were taken from boreholes representing the shallow alluvial aquifer of Ouled Djellal (Fig. 1). Approximately 1.5 dm³ polyethylene bottles have been used to store non-acidified water. Next, samples were kept at 4 °C and their chemical composition was determined by studying the components of Ca^{+ 2}, Na⁺, K⁺, Mg^{+ 2}, HCO₃^-, SO₄^2-, and NO₃^-, and Cl^-. The standard analytical methods were used to analyze water at the hydrochemistry laboratory of Constantine University in Algeria. Titration method with the methyl orange end point was conducted for the determination of the bicarbonates. The chloride level was estimated using the Mohr technique of titrating and precipitating AgCl until silver chromate appears³⁹. The nephelometric technique was used to determine sulfate⁴⁰, while colorimetric analysis was used to examine nitrates⁴¹. A complex metric titration was used to determine the amounts of calcium and magnesium. To measure the sodium and potassium concentrations, emission spectrometry was utilized³⁹.

Twenty-three groundwater samples were collected from electrically pumped boreholes at the different locations in the study area, representing the shallow alluvial aquifer of Ouled Djellal (Fig. 1). Following a 15-minute pumping period to remove any stored groundwater, samples were taken using two acid-washed polypropylene (PP) bottles, following the American Public Health Association’s protocol⁴². Using cellulose acetate 0.45 m filters, each sample was quickly filtered locally. Before being acidified to a pH of 2 with 5 mL of 6 N HNO₃, the filtrate was immediately transferred to 1.5 L polyethylene (PE) bottles for the cation analysis. Without acidification, samples were transferred to 1.5-litre plastic bottles for anion analysis. Prior to being transported to the lab for analysis, our samples were maintained at a temperature of around 4 °C in an ice chest. Major ions analyzed included Ca²⁺, Mg²⁺, Na⁺, K⁺, HCO₃⁻, SO₄²⁻, Cl⁻ and NO₃⁻, with all concentrations expressed in mg/L. At the location, the physicochemical parameters (electric conductivity, temperature, TDS, and pH) were measured with a WTW multiparameter (P3 MultiLine pH/LF-SET) made in Germany. With its conductivity cell and pH/redox electrode connected in parallel, the WTW Multi-Line P3/LF-SET multi-parameter meter can measure all three parameters simultaneously. It also has a linear and non-linear temperature compensation function, so it can be used with both natural and ultra-pure waters.

Chemical analyses were carried out at the Water Control and Quality Laboratory of the Biskra Unit of the Algerian Water Company (ADE Biskra), Algeria. Bicarbonate concentrations were determined in the field after sampling by titration with H₂SO₄ using methyl orange as the endpoint. In the laboratory, chloride was measured by the Mohr method (argentometric titration until the appearance of silver chromate). Sulphate was quantified using the nephelometric method, while nitrate was analysed by colorimetric analysis. Calcium and magnesium concentrations were determined by complexometric titration with EDTA. Sodium and potassium were measured by flame emission spectrometry. The TDS was estimated by weighing and drying at 103–105 °C in an oven. This was achieved by applying standard methods of the APHA 2017. Quality assurance included duplicate analyses of 10% of samples, regular calibration with certified reference standards, and blank analyses. To check the correctness of the analysis, the ionic balance was used with charge balance errors maintained within ± 5%, ensuring data reliability and analytical accuracy.

Multivariate statistical techniques and data analysis

Cluster analysis (CA)

It is a technique that identifies the distinctive characteristics of each group by grouping enormous datasets from each entity into various clusters^43,44,45. It is commonly used to classify hydrogeochemical processes in Groundwater by dividing collected water samples into important geological and hydrogeological categories, particularly for hydrochemistry investigations^43,46. The cluster dendrogram was applied to offer a visual representation of the clustering activities by greatly showing the groupings and their proximity while reducing the complexity of the original data.

Principal component analysis (PCA)

It is known as a linear structure with complex multivariate datasets methods that may be efficiently statistically analyzed without scarifying information⁴⁷. The quantity of variables can be reduced while maintaining the same degree of related variability⁴⁸. At the end, the PCA remains an effective tool for a better understanding of the relationships between information about basic, indirectly observable features.

Indexing approach

Chloro-alkaline indices

CAI-I and CAI-II were used to show the relationship and the origin and between the major elements such as SO₄^2- versus (Ca²⁺; Mg²⁺), (HCO₃^-+ SO₄^2-) versus (Ca²⁺+Mg²⁺), (Mg²⁺/Ca²⁺) versus (Mg²⁺/Na⁺), (HCO₃^-) versus (Ca²⁺+Mg²⁺). They were applied to determine the ion exchange mechanism between GW and aquifer minerals (Eqs. 1 and 2)⁴⁹.

$$\:\text{C}\text{A}\text{I}-\text{I}=\frac{{\text{C}\text{l}}^{-}-\:({\text{N}\text{a}}^{+}+{\text{K}}^{+})}{{\text{C}\text{l}}^{-}}$$

(1)

$$\:\text{C}\text{A}\text{I}-\text{I}\text{I}=\frac{{\text{C}\text{l}}^{-}-\:\left({\text{N}\text{a}}^{+}+{\text{K}}^{+}\right)}{{{\text{S}\text{O}}_{4}}^{2-}+{{\text{H}\text{C}\text{O}}_{3}}^{-}+{{\text{C}\text{O}}_{3}}^{2-}+{{\text{N}\text{O}}_{3}}^{2-}}$$

(2)

Irrigation water quality indices

In this study, six indexes including IWQI, SAR, PS, SSP, RSC, and Na%, respectively, were calculated using the physicochemical parameters of the groundwater samples as given in Table 2.

Table 2 The IWQIs, formula, and its reference.

Irrigation water quality index

Equation 3 was used to apply a non-dimensional scale with a range of 0–100 to the relationship between variables (Na⁺, EC, SAR, Cl, and HCO₃⁻), in order to calculate IWQI:

$$\:IWQI=\:\sum\limits_{i=1}^{n}{Q}_{i}\times\:{W}_{i}$$

(3)

Q_i indicates the quality measurement’s results within the tolerance ranges, and W_i indicates the weight of each parameter (as seen in Table 3).

$$\:{Q}_{i}={Q}_{max}-\left(\frac{\left[\left({X}_{ij}-{X}_{inf}\right)\times\:{Q}_{imap}\right]}{{X}_{amp}}\right)$$

(4)

where $\:{X}_{inf}$: correspond to the lower limit of the class, X_ij: is the observed value (for each parameter), X_amp: is the amplitude class that the parameter falls within, and Q_imap is the class amplitude. Equation 5 was used to obtain the values of W_i:

$$\:{W}_{i}=\frac{\sum\:_{j=1}^{k}{F}_{j}{A}_{ij}}{\sum\:_{j=1}^{k}\sum\:_{i=1}^{n}{F}_{j}{A}_{ij}}$$

(5)

where $\:i$ is the number of physicochemical parameters chosen by the model (from 1 to n), $\:\text{F}$ is the auto value of component 1, $\:\text{j}$ is the number of selected factors by the model (from 1 to k). Finally, $\:\text{A}$ = The substantially limited of parameter $\:\text{i}$ by factor $\:\text{j}$.

Table 3 The range of limit values of the parameters used in the computation of quality measurement.

Datasets and artificial intelligence (AI) s approaches

In this study, model training and validation were carried out using representative data samples to ensure reliable performance evaluation. The dataset was randomly divided into 70% (training set) and 30% (validation set), which is a widely adopted practice in water quality modeling. As an example, some researchers employed a similar data split when training Deep Learning (DL) models for water quality index prediction²³, reaching consistent performance across multiple evaluation metrics (such as R² and RMSE). In the same vein²⁴, implemented random data partitioning to assess the generalization ability of multilayer perceptron (MLP) networks and regression models in WQI. There are also other recent studies, such as^32,55 who have also adopted comparable ratios (like 70–30, 80–20, or 75–25) to maintain an effective balance between adequate training data and validation accuracy.

In our case, three models: multiple linear regression (MLR), artificial neural network (ANN), and support vector machine (SVM), were trained and evaluated (Fig. 4). We adopted this procedure in order to reduce the risk of overfitting and to ensure a reliable assessment of each model’s predictive performance. All data preparation, modeling, and analysis tasks were performed using SAS JMP Pro 18 software. The computations were carried out on a computer equipped with an Intel Core i7-3630 QM CPU (2.4 GHz) and a RAM of 8 GB.

Fig. 4

Flowchart for prediction of the IWQs using MLR, ANN, and SVM.

Artificial neural networks (ANNs)

Artificial intelligence (AI) approaches have recently been acknowledged as powerful methods for modeling complex nonlinear phenomena in hydrology⁵⁶. As an example, since 1980 s, ANNs have emerged as highly effective tools specifically to predict water quality⁵⁷. These models consist of interconnected layers, input, hidden, and output with non-linear activation functions such as logistic, tanh, or ReLU, which allow them to uncover hidden patterns in water data without depending on predefined mechanistic rules^33,58,59,60. A recent review indicated that ANN architectures are well-suited to a wide range of aquatic systems, from rivers and lakes to reservoirs and wastewater treatment plants, and excel at forecasting diverse water quality parameters⁶¹. Generally these models achieve a high predictive accuracy thank to their adaptability and minimal requirement for a priori assumptions, ANNs stand out as robust approaches for continuous monitoring and management of the quality of water⁶².

Support vector machine (SVM)

SVM models are a highly effective ML technique, specifically for assessing water quality, due to their robust generalization ability and resilience against overfitting^63,64. SVMs are valuable for both classification and regression tasks, especially in environmental science. Some studies have demonstrated that SVMs can predict key water quality parameters with high accuracy⁶⁵, while another study presented the fact that the SVM model accurately estimate the water quality factors with minor prediction errors, confirming its efficacy and performance⁶⁶. In addition, hybrid models integrating SVM with WQI have shown strong predictive performance, achieving high classification accuracies as it was the case in groundwater quality assessment of the Egyptian Delta⁶⁷, same thing for⁶⁸, where SVM was also the most accurate algorithm for water quality classification prediction in India. Thus, SVM remains a good choice and has excellent advantages in dealing with small samples and complex nonlinear model problems. In comparison with other algorithms, it has advantages such as fast learning speed and strong generalization ability^69,70.

Multiple linear regression (MLR)

MLR remains a key method across various fields due to its computational efficiency, interpretability and also its well-established theoretical framework⁷¹. Its capacity to quantify relationships between a dependent variable and multiple predictors facilitates robust predictive modeling while controlling for confounding factors. Recent studies still use MLR, such as⁷²who added a brute-force approach⁷³. also used MLR to evaluate water quality in Hanyuan Lake (China). In recent years, with advancements in machine learning (ML) models, MLR appear to be less frequently employed in some fields, in favor of methods such as ANN, SVM, random forests (RF), or XGBoost. Furthermore, the results concerning the efficacy of MLR are mixed, for example⁷⁴, found that MLR were less accurate than ANN in predicting water quality parameters in Iran, while⁷⁵ and⁷⁶ demonstrated that both MLR and ML models were very close and had a high accuracy in water quality evaluation in Syria and Turkey, respectively.

Evaluation of model performance and sensitivity analysis

To investigate the effect of model architecture on predictive accuracy, hyperparameter sensitivity analyses were conducted for three models (SVM, ANN, and MLR). For the SVM model, a linear kernel was applied, and the cost parameter (C) was varied between 0.01 and 5.0 across 20 independent trials. The optimal configuration was determined using the highest coefficient of determination (R²), the lowest Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) values. While the ANN architecture comprised between one (1) and three (3) hidden layers incorporating multiple activation functions (tanh, sigmoid, and linear), to enhance nonlinear learning and generalization capacity. The tanh and sigmoid functions were primarily employed in the hidden layers to capture complex nonlinear relationships, while a linear activation function was used in the output layer to generate continuous index values. The network was trained with a learning rate of 0.1, using 12 input features (physicochemical parameters) and a single output neuron corresponding to the target variable (indexes such as SAR, IWQI, SSP, Na%, RSC, or PS). Model weights were optimized through the backpropagation algorithm to minimize the Mean Squared Error (MSE). This structure mirrors practices seen in recent WQI studies, such as^23,24 where model hyperparameters were systematically tuned to enhance model robustness and generalization. The efficiency of the regression model was calculated using the statistical metrics root mean square error ($\:RMSE$), according to formula n° 6 and coefficient of determination ($\:{R}^{2}$), according to formula n° 7. Both indices are calculated as follow^75,76:

Root Mean Square Error (RMSE)

$$\:RMSE=\:\sqrt{\frac{1}{N}{\sum\:}_{i=1}^{N}{({F}_{\:act}\:-\:{F}_{\:p})}^{2}\:}$$

(6)

Determination Coefficient (R²)

$$\:{R}^{2}=\:\frac{\sum\:{({F}_{\:act}\:-\:{F}_{\:p})}^{2}}{\sum\:{({F}_{\:act}\:-\:{F}_{\:ave})}^{2}}$$

(7)

It is possible to explain all parameters by: F_p which is the calculated or the simulated value, N stands for data entries total number, while $\:{F}_{\:act}$ stands for the actual real value that is based on laboratory analysis, and finally, F_ave stands for the mean value.

Data processing, analysis, and Spatial mapping

The software Statistical Package for the Social Sciences “SPSS” (version 22) has been employed to perform statistical analyses on IWQIs and the physicochemical characteristics. Piper’s diagram was utilized to determine water types and assess hydrogeochemical evolution based on the composition of cations and anions. The PCA and CA were applied using the software Statistica (v.8) to determine the key constituents of GW and to enhance water quality assessments by simplifying data analysis into recognizable patterns. Then, spatial distribution maps of the groundwater quality indices were generated using the Inverse Distance Weighted (IDW) interpolation and Empirical Bayesian Regression Kriging (EBRK) methods using ArcGIS Pro (v 3.0.1). In order to assess the reliability of the interpolation results and minimize potential bias arising from the spatial distribution of sampling wells, cross-validation was performed using the leave-one-out approach. The validation included the computation of statistical performance indicators such as the Mean Error (ME), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R²). These metrics were used to compare the performance of IDW² and EBRK, where lower RMSE and MAE and higher R² values indicated better model accuracy and prediction reliability. Finally, Multiple Linear Regression (MLR), Artificial Neural Network (ANN), and Support Vector Machine (SVM) models were computed using JMP 18 Pro (JMP, Cary, NC, USA).

Results and discussion

Hydrochemical properties of groundwater

Twenty-three GW samples from Ouled Djellal area were collected and subjected to analysis. They were studied experimentally with chemometric techniques. Water quality index and statistical methods like correlation analysis, principal component analysis, cluster analysis, and GIS techniques were used to study the hydrochemical parameters. Table 4 showed a descriptive summary of the parameters of the analyzed groundwater samples, which presented alongside the Food and Agriculture Organization (FAO) standards⁷⁷. Below, the spatial patterns of major ionic constituents (Ca²⁺, Mg²⁺, Na⁺, K⁺, HCO₃⁻, SO₄²⁻, Cl⁻, NO₃⁻) in the study area’s groundwater samples are discussed. The statistical analysis of the twenty-three groundwater samples (Table 4) shows that the hydrochemical parameters vary a lot. This is because of both natural geochemical processes and human activities. For several major parameters recommended by the FAO, the levels are well above the standards. The total hardness (TH) reaches very high levels (mean of 4950 °F), confirming the high levels of calcium and magnesium ions.

Table 4 Summary of descriptive results of groundwater samples in the study area.

In general, the distribution of calcium (Ca²⁺) showed a marked concentration in the northern part of the region, as indicated by sample 20, which presented darker tones on the spatial distribution map. These high concentrations (ranging from 232 to 701 mg/l, with an average of 469 mg/l) frequently exceed the FAO limit (400 mg/L) and primarily reflected the dissolution of carbonate minerals (calcite, dolomite) and evaporitic minerals (gypsum). For magnesium (Mg²⁺), it shows a distribution similar to calcium, with higher concentrations in the northern part of the plain. The concentrations range from 75 to 780 mg/l, with an average of 194.65 mg/l, well above the FAO standards (60 mg/l). This distribution reflected the alteration of magnesium minerals and cation exchange processes, particularly active in the clay soils of the Ouled Djellal region. As for the distribution of sodium (Na⁺), it revealed maximum concentrations localized in the center-north and at certain points in the center of the region, as illustrated by sample 20 with dark tones. The concentrations varied from 60 to 1002 mg/l, with an average of 254.74 mg/l, generally not exceeding the FAO threshold (920 mg/L), with a high coefficient of variation (R = 0.9) indicated strong spatial heterogeneity. This distribution reflected the accumulation of soluble salts through evaporation and the possible anthropogenic inputs related to irrigation return flows. Finally, potassium (K⁺) follows a distribution similar to sodium, with relatively high levels in the same geographical areas, although in lower concentrations (5 to 70 mg/l, with an average of 21.57), which clearly exceed the recommended value (2 mg/L), as shown in Table 4. This distribution indicated a mixed origin between silicate weathering and agricultural inputs. The analysis of the spatial distribution of cations in the Ouled Djellal aquifer revealed a strong geographical heterogeneity due to evaporation processes, which characterize arid regions^36,78,79,80.

For the geographical distribution of anions, chlorides (Cl⁻) showed the highest concentrations in the northeast of the region, with extremely variable values (425 to 4467 mg/l, with an average of 1126.52 mg/L) exceeding the FAO recommended limit of 1065 mg/L, as seen in Table 4. As an example, sample 20 reflected a high concentration of Cl⁻. This distribution reflected the intense evaporation process, as well as possible anthropogenic influences (irrigation returns flow and domestic discharges). Sulfates (SO₄²⁻) concentrations presented a significant quantities in the center of the region, which ranging from 207 to 560 mg/l and remaining below the FAO threshold (960 mg/L). This distribution is primarily due to the dissolution of gypsum in geological formations. Moreover, samples 21 and 18 suggest the highest concentrations of SO₄²⁻. Bicarbonates (HCO₃⁻) showed high levels in the north and northeast, which ranging from 195 to 800 mg/l in some samples, which exceed the FAO threshold value (610 mg/L), that confirm the significant influence of carbonate dissolution processes in hydrochemical evolution. Nitrates (NO₃⁻) values varied from 2.78 to 87 mg/L (average 28.04 mg/L) and significantly exceed the FAO standard (10 mg/L) at several points, due to nitrogen fertilizers and agricultural effluents^81,82.

The pH values revealed a remarkably homogeneous across the entire region (7.16 to 7.86 with an average of 7.48), that indicated a neutral to slightly alkaline environment, in line with the range recommended by the FAO for irrigation (6.0–8.5) and consistent with the carbonated nature of the geological substrate⁸¹. The EC values reached to 16,100 µS/cm with an average of about 6,995 µS/cm and the TDS reached to 8,050 mg/L with an average of about 3,498 mg/L, exceeding the FAO thresholds (3000 µS/cm and 2000 mg/L, respectively), which indicated a high mineralization that could affect the quality for agricultural use by posing salinity problems for soils and crops. This spatial distribution perfectly coincides with the accumulation zones of major cations and anions. The spatial analysis showed that the distribution of hydrochemical parameters in the shallow aquifer of Ouled Djellal is primarily controlled by evaporation processes in an arid context, with a preferential concentration of salts in the northern and central areas of the region. This spatial heterogeneity, resulting from the complex interaction between climatic, geological, and anthropogenic factors.

Hydrogeochemical characteristics and water types

It is well established that GW chemistry is largely influenced by the type of rock formations that flows through and the duration of contact with those layers. As a result, the elements found in the water offer valuable insight into the characteristics of the aquifer that passes through^12,33. Hydrochemical facies roughly assess the quantities of water that differ in their chemical composition, which is a combination of water-rock interactions through the rock matrix, solution kinetics, hydrogeological and environmental contexts, which is determined by calculating reaction quantities, considering ionic composition. Among the most widely used graphical methods for determining groundwater types are Piper and Gibbs diagrams^83,84. The first one is a trilinear diagram that comprises a diamond and two equilateral triangles representing anions and cations respectively. Here, the diagram was constructed using Diagrams V 8.6 software. A single type of groundwater in Ouled Djellal was identified and classified according to their position on the diagram of Piper (Fig. 5). This group depended on the lithology of the area and the hydrodynamics of the water table. The cations fall within zones B and D, suggesting that the GW is mainly of the calcium type, with some samples showing no clear dominant cation. As for the anions, most of the samples are predominantly located in zone G, indicating a high prevalence of chloride. In the resulting diagram, all our samples are positioned in the upper zone, with Cl⁻, Ca²⁺ and Mg²⁺ being the main ions. An analysis of the molar concentrations of the different elements in the research area reveals the following trend among the cations: Ca²⁺> Mg²⁺> Na⁺> K⁺ evolving as follows: Cl⁻> SO₄²⁻> HCO₃⁻> NO₃⁻. Thus, the groundwater type in Ouled Djellal aquifer revealed chloride water type for anions, and calcium and magnesium water type for cations, which classified as calcium-magnesium chloride water types due to the presence of saline, sulfate-rich evaporite formations, notably gypsiferous clays. The main ions distribution (Ca²⁺, Mg²⁺, Cl⁻ and SO₄²⁻) is highly influenced by the lithology of the region and by anthropogenic activity, which leads us to think of an anthropogenic origin, like the quality of irrigation water, uncontrolled fertilization and domestic discharges from urban clusters dotted around this region.

Fig. 5

Hydrogeochemical characteristics and controlling mechanisms for the collected groundwater samples. (a) Piper diagram and (b) Gibbs diagram.

Processes influencing groundwater chemistry

The chemical composition of GW samples indicate that the impact of hydrochemical processes in the aquifer including evaporation and rock-water interaction^36,85,86. Statistical analysis of the major ions (K⁺, Ca²⁺, Na⁺, Mg²⁺, SO₄²⁻, HCO₃⁻, Cl⁻, and NO₃⁻) provides the fundamental basis for comprehending groundwater hydrogeochemical properties, which showed concentrations above the permissible limit for TDS and HT. Plotting of total dissolved salts (TDS) Vs. Total Hardness (TH). Figure 6 showed that 47% of GW samples were soft-brackish water and about 53% of samples were hard-brackish water. The interrelationships between anions and cations are a useful tool for governing mechanisms of GW chemistry.

Fig. 6

Plot of TDS Vs TH of groundwater samples.

Evaporation, soil-salt leaching and rock weathering

Gibbs diagram was used to interpret the changes in GW chemistry in the region of Ouled Djellal⁸⁴. It classifies evolutionary processes into three types according to the dominance of rocks, precipitation and evaporation (Fig. 5). In our case, only one groundwater type was identified (soft-brackish water) in the study area. This group appears in the upper section of the diagrams, pointing to an evaporative origin and suggesting that evaporation is determining in shaping groundwater chemistry, while rock weathering has a minor influence. In fact, the Ouled Djellal area belongs has an arid climate with a high evaporation rate, which can explain the diversion of the water points towards the evaporation dominance zone (Fig. 5). This result is corroborated by the projection results on the Piper diagram, which indicating the importance of lithology on the chemical facies of GW in the studied region.

Based on Gibbs diagram, the ratio Ca²⁺/Na⁺ served as an indicator of salt leaching from soils. Nonetheless, the ratio Ca²⁺/Na⁺ may not be reliable in desert environments, as high evaporation rates and the abundance of sodium salts, commonly found in desert soils, which can elevate sodium levels relative to calcium due to water–rock and water–soil interactions^87,88. As a result, Mg concentrations are not significantly affected by salt leaching from the soil or by saline precipitation during the early stages of evaporation. This makes Mg-to-cation ratios useful indicators of key processes in desert environments, especially the Mg²⁺/Na⁺ ratio for identifying salt leaching and the Mg²⁺/Ca²⁺ ratio for tracking evaporation. In this study, the Mg/cation plot (Fig. 7a), consistent with the Gibbs diagram, showed that the samples exhibit elevated Mg/Na and Mg/Ca ratios. These results suggest that evaporation was a key determinant in shaping groundwater chemistry, while water–rock interactions are likely a secondary factor. Our GW samples fall within the evaporation zone of the Gibbs diagrams, which indicating that the main components of the groundwater in the study area affected by intense evaporation, which leads to high TDS concentrations. In addition, the Gaillardet diagram⁸⁹ was utilized to capture various hydrochemical reactions in unmixed conditions. The observed correlation (R = 0.509) between Na⁺ and Ca²⁺ indicated an exchange process between these ions in the groundwater flow⁹⁰ (Fig. 7b). Therefore, the collected groundwater samples in Ouled Djellal aquifer were affected by silicates weathering and evaporation dissolution processes. The strong correlation between Ca²⁺/Na⁺ Vs. HCO₃⁻/Na⁺, R = 0.97 and Ca²⁺/Na⁺ Vs. Mg²⁺/Na⁺, R = 0.83 in log-log space. Based on Fig. 7b, the composition of the GW is mainly governed by evaporation dissolution and silicate weathering.

The ratio Mg²⁺/Ca²⁺ indicated the dolomite and calcite dissolution processes. Mg²⁺/Ca²⁺ ratio > 1 also indicated the dolomite dissolution, while a ratio less than unity indicated the calcite dissolution^91,92 in clay soils, due to ion exchanges between clay minerals and the solution, which lead to an enrichment of the water in calcium and released or adsorbed sodium according to local equilibrium conditions. The ion exchange processes can increase Ca²⁺ concentration with a proportional decrease of Na⁺⁹³. In this study, GW samples have a ratio < 1, that reflecting the rapid dissolution of calcite relative to dolomite (Fig. 7c), and the predominance of Ca²⁺ in GW appears to derive from carbonate and evaporite mineral dissolution, with gypsum dissolution contributing as a secondary input.

Silicate formations are present in the clayey formation and siliceous sands, which widely found in the studied region of Ouled Djellal in the beds of wadis and the banks of temporary wadis. Their alteration and interaction with groundwater promote the gradual release of Na⁺ and Mg²⁺ ions^94,95. A ratio HCO₃⁻/Na⁺ > 1 suggested carbonates weathering, while reduced levels imply silicate-weathering dominance. Plotting the groundwater samples on the scatter plot of HCO₃⁻ versus Na⁺ and in relation to slope line 1 (Fig. 7d), denoting a slight majority (52%) of the GW, which indicated silicates weathering processes. The scatter plot of (HCO₃⁻ + SO₄²⁻) versus (Ca²⁺ + Mg²⁺) was far above the slope line 1 (Fig. 7e), that highlighting the dominance of magnesium and calcium containing minerals. The observed mineralization can be linked to the weathering and dissolution of carbonate minerals, along with contributions from other sources of Mg²⁺ and Ca²⁺, such as gypsum or anorthite, or from changes due to cation exchange processes^91,96.

Ions exchange processes

For understanding the hydrochemical reaction affecting on the groundwater quality, ion exchange in Eq. 8 and reverse ion exchange in Eq. 9 were applied⁹⁷. These methods were used to determine the occurrence of cations exchange process, supported withthe relation between (Ca²⁺ + Mg²⁺ - HCO₃⁻ - SO₄²⁻) Vs. (Na⁺ + K⁺ - Cl⁻)⁸⁹. If cation exchange significantly influences the ionic composition of groundwater, the relationship between these parameters is expected to be linear, with a slope of Y = –X. In this study, a strong linear alignment with the reference trend Y=-X. Figure 7f indicated an increase in Ca²⁺ + Mg²⁺ linked to a decrease in Na⁺ + K⁺ or probably an increase in HCO₃⁻ + SO₄²⁺. The strong correlation R = 0.98 between Ca²⁺ + Mg²⁺ - HCO₃⁻ - SO₄²⁻ & Na⁺ + K⁺ - Cl⁻, indicated that cations play a key role in regulating the hydrochemical composition of groundwater and the involvement of Ca²⁺, Mg²⁺, and Na⁺ in ion exchange reactions⁹⁸, as represented by Eq. 8.

To analyze the ion exchange process, a plot of Na⁺ against Cl⁻ was used. Theoretically, halite dissolution should release equal amounts of Na⁺ and Cl⁻ as displayed in Eq. 10, and the concentrations of these ions in meteoric water infiltrating into groundwater are expected to be nearly equal. Therefore, halite dissolution regulates the concentrations of Na⁺ and Cl⁻ in the groundwater system, whereas sodium released from silicate weathering results in a broader Na⁺/Cl⁻ ratio^36,99. Moreover, the scatter plot Na⁺ versus Cl⁻ in all samples was under the line of halite dissolution, indicate an excess of Cl⁻, due to other Cl origin (Fig. 7g). Finally, the elevated concentration of Cl⁻ in GW is mainly attributed to climatic factors like intense evaporation, but it may also result from anthropogenic sources like domestic wastewater and irrigation return flow.

$$\:2NaX+{Ca}^{2+}\to\:2{Na}^{+}+{CaX}_{2}$$

(8)

$$\:Ca{X}_{2}+2{Na}^{+}\to\:{Ca}^{2+}+2NaX$$

(9)

$$\:NaCl\to\:{Na}^{+}+{Cl}^{-}$$

(10)

Anthropogenic input

Chloride behavior conservatively in natural waters, as it is largely unaffected by chemical, biological, or physical processes. The NO₃⁻/Cl⁻ ratio is a useful indicator of mixing processes or biological activity affecting nitrate and chloride in groundwater. Chemical fertilizers typically contain high levels of nitrogen and low amounts of chloride, whereas domestic and animal wastewaters tend to have high chloride concentrations and low NO₃⁻/Cl⁻ ratios¹⁰⁰.

Figure 7h showed the variations in the NO₃⁻/Cl⁻ molar ratio as a function molar concentration of Cl⁻ in groundwater and different poles and potential vectors of nitrates confirm the agricultural origin for nitrates and domestic wastewater or the return of irrigation water for chlorides. The low correlation coefficient (R = 0.38) between NO₃⁻/Cl⁻ and Cl⁻ concentration testifies to a different exogenous origin of nitrates and chlorides.

Fig. 7

Stoichiometric relationships between major cations and anions of the groundwater samples in Ouled Djellal aquifer. (a) (Mg/Na and Mg/Ca), (b) ((Ca/Na) vs. (HCO3/Na) and (Ca/Na) vs. (Mg/Na)), (c) (Na Vs. Mg), (d) (HCO₃ Vs. Mg), (e) ((HCO₃ + SO₄) Vs. (Ca + Mg)), (f) ((Ca + Mg - HCO₃ - SO₄) Vs. (Na + K - Cl)), (g) (Na Vs. Cl), (h) (NO₃/Cl Vs. Cl).

Chloro-alkaline index

During groundwater flow, the water is exposed to certain formations that have the properties of exchanging their ions for those continuous in the water. The basic exchange index is the ratio between the ions exchanged and the ions of the same nature originally existing. Two basic exchange possibilities can occur: either the exchange of alkalis with alkaline earths (CAI-I), according to Eq. 1, or the exchange of alkaline earths in water with alkalis (CAI-II), according to Eq. 2. Generally, K+ & Na + in GW are exchanged with Ca²⁺ & Mg²⁺ in aquifer matrix, resulting in a direct ion exchange¹⁰¹. The GW samples showed a positive value for both indices (CAI-I & CAI-II) suggesting the contribution of direct ion exchange in the system (Fig. 8).

Fig. 8

Cations exchange diagram for the groundwater in Ouled Djellal aquifer.

Geochemical modelling

Geochemical modelling was used to explain the mechanisms by which mineralization occurs. This was carried out by simulating natural evaporation using the Phreeqci software¹⁰², which computes aqueous speciation and thermodynamic equilibrium conditions of groundwater, considering the dominant mineral phases in the aquifer. Mineral reactions (dissolution or precipitation) are constrained by their respective saturation indices. Calculation of the saturation index (SI) of dissolved minerals in water has been carried out using Debye & Hückel’s law, as explained by¹⁰³ and expressed as follows Eq. 11.

$$\:IS=\:\frac{Log\:PAI}{k}$$

(11)

when, IS > 0.5, water is supersaturated with respect to a mineral. This mineral would tend to precipitate, against when IS < −0.5, the water is undersaturated with respect to a mineral. This indicates the minerals tendency to dissolve.

The calculated values of the SI for carbonates and evaporates elements are depicted in Table 5. The SI values were calculated for the carbonate elements such as, calcite, aragonite, and dolomite (Fig. 9). When the SI > 0.5, demonstrating that the GW was supersaturated with regard to these minerals, and precipitation could have occurred. The supersaturation of these elements may lead to a preferential precipitation, which could diminish the concentrations of Ca²⁺, Mg²⁺ and HCO₃⁻ in groundwater due to high temperature, high evaporation rate, and low rainfall. On the other hand, the SI < 0.5 for gypsum, anhydrite, and halite, suggesting persistent under-saturation. Consequently, Na⁺, Cl⁻, SO₄²⁻, and Ca²⁺ ions may continue to be released into groundwater through weathering of these minerals. The dissolution of these minerals was indicated by the high Na⁺ & Cl⁻ correlation (R = 0.95) and Ca²⁺ & SO₄²⁻ correlation (R = 0.43).

Table 5 Descriptive results of the calculated SI in the study area.

Fig. 9

Saturation index for carbonates and evaporates in Ouled Djellal aquifer.

Statistical analysis

Elementary statistics and correlation analysis

The correlation matrix (Fig. 10) for the variables revealed a close relationship between the various elements, such as Na⁺, Cl⁻, Mg²⁺ K⁺, Ca²⁺ and HCO₃⁻ which are significantly correlated with EC and TDS (0.502 < R < 0.894). These correlations indicated that water chemistry in the Ouled Djellal region controlled by lithological factors and various hydrogeochemical processes. These results demonstrate a strong influence of these elements on regional groundwater chemistry. Conversely, the lack of significant correlation between EC/TDS and SO₄²⁻/NO₃⁻ implies negligible involvement of these anions in salinization processes.

Significant positive correlation values were observed between all the elements except NO₃⁻ and SO₄²⁻. These trends resulting from leaching processes and dissolution of evaporite minerals. While chloride salts may release Cl⁻ into GW, while the high Na⁺ - Cl⁻ (0.954) and K⁺ - Cl⁻ (0.775) primarily indicated dissolution of evaporitic salts such as halite due to the influence of evaporation. A secondary contribution may arise from silicate alteration, which can affect the mobility of Na⁺ and K⁺ associated with Cl⁻ in certain lithological contexts³⁶. According to the SO₄²⁻ concentrations results (max 11.67 meq/L^{− 1}) at borehole 18, the origin can be attributed to the limited dissolution of sulphur minerals (a trace of gypsum or locally oxidized pyrite), which found in limestone, sand. While the element NO₃⁻ showed an insignificant negative correlation with the whole elements, which indicated an anthropogenic origin including agricultural practices. Strong correlations underline dissolution of gypsum, halite, and salts as the primary solute source, exception for HCO3⁻ (carbonate weathering) and NO₃⁻ (anthropogenic input). Figure 10 indicated that R value for Mg²⁺, Cl⁻, K⁺ and NO₃⁻ in Ouled Djellal area was relatively high, due to more reactive of ions to hydrological, geological, and anthropic factors. The statistical results are fully consistent with the geochemical evidence, confirming that evaporation and saline mineral dissolution were the dominant mechanisms shaping GW composition in the shallow Oued Djellal aquifer. Minor discrepancies among certain ions (SO₄²⁻ and NO₃⁻) are clearly explained as resulting from localized geological or anthropogenic inputs.

Fig. 10

Scatterplot and correlation matrix for the groundwater in the Ouled Djellal aquifer.

Principal component analysis (PCA)

According to the PCA results, F1 accounts for 62.82% of the total variance, indicated a well-defined structure in the sampled data, which suggests a strong relationship among the key factors influencing sample composition. These relationships were further supported by the correlation matrix (Fig. 10), which revealed significant positive correlations between Na⁺, Cl⁻, Mg²⁺, K⁺, Ca²⁺, HCO₃⁻, EC, and TDS (r = 0.89, 0.87, 0.87, 0.70, 0.52, and 0.50 (Fig. 10). Analysis of F1-F2 (Fig. 11a) showed that the F1 factor is strongly positive loading of EC, TDS, Na⁺, Cl⁻, Mg²⁺ and K⁺ and to a lesser extent by Ca²⁺, demonstrating that salinity originates from multiple sources, including evaporite dissolution, leaching of saline soils (containing clays, gypsum, and anhydrite), aqueous ion exchange processes, and anthropogenic contributions (Table 6).

At the same time, F2 explains 13.01% of the total variance of the data set and demonstrates lower positive loadings of SO₄²⁻ and HCO₃⁻ (0.385 and 0.329), respectively (Fig. 11a). These results indicated that SO₄²⁻ and HCO₃⁻not influence on water mineralization and predicts the different lithological origin. This likely results from carbonate weathering, which reflecting the impact of acid-base equilibrium on groundwater chemistry¹⁰⁴ coupled with dissolution of sulfate minerals (notably gypsum) from the soil matrix. Regarding F1-F3, F3 explains 10.23% (Fig. 11b) and revealed a significant positive NO₃⁻ load (0.869). Table 6 indicated the arrival of agricultural pollution or a mixture of deep water with surface water. This factor is associated with chemical inputs to the farm, mineralization of soils and non-agricultural sources or deep water mixing with surface water. Oxidation conditions promote the microbial oxidation of fertilizer NH₄⁺ to NO₃⁻ through nitrification processes¹⁰⁵, as shown in the following reaction (Eq. 12).

$$\:{NH}_{4}^{+}\:+\:2{O}_{2}=\:{NO}_{3}\:+\:{H}_{2}O\:+\:2{H}^{+}$$

(12)

Table 6 Factor analysis of the variables.

Fig. 11

PCA scores: (a) F1 vs. F2 (b) and F1 vs. F3.

Cluster analysis (CA)

Hierarchical Ascending Classification (HAC) was applied using City-block (Manhattan) as the measurement distance between the analyzed GW samples. Ward’s technique was used as the linking rule for the classification of data acquired from water in Ouled Djellal. The results are presented in the form of a dendrogram of the 10 parameters, such as EC, TDS, Mg²⁺, K⁺, Ca²⁺, Na⁺, HCO₃⁻, SO₄²⁻, Cl⁻, and NO₃⁻ (Fig. 12). Analysis of the diagram revealed three distinct statistical classes and indicated that TDS and EC may serve as key discriminators among elemental parameters (Fig. 12).

The first group indicated a strong correlation between Mg²⁺, Cl⁻, Na⁺ K⁺, EC and TDS, emphasizing the dominance of evaporitic elements in the GW chemistry of Ouled Djellal aquifer (Fig. 12). This evaporitic signature revealed Cl⁻- Ca²⁺ - Mg²⁺ facies. In contrast, Group 1 revealed a weak Mg²⁺-Ca²⁺ correlation, which indicative of Ca²⁺ dual origin from both carbonate and evaporitic sources. The second group showed a close association between NO₃⁻ and SO₄²⁻ and a dissociation of the other chemical elements present in the GW of Ouled Djellal aquifer, due to the anthropogenic origin of NO₃⁻ and, lithological through the dissolution of gypsiferous clays and anthropogenic for SO₄²⁻. Group 3 revealed an association between HCO₃⁻ and Ca²⁺, which reflecting the same carbonate origin.

Fig. 12

Cluster dendrogram for variables.

Suitability of water for irrigation

More than 60% of the samples from the shallow Ouled Djallal aquifer showed a moderate to high risk of salinization associated with EC values exceeding 3000 µS/cm (Fig. 13a), which indicating a danger of soil alkalization and lead to clay dispersion and decrease permeability. The Wilcox diagram, also based on SAR and EC showed that most of the groundwater samples fall into the C3–C4/S1–S2 classes, which indicating excessive salinity (C4) (Fig. 13b) due to water-rock interactions, the dissolution of gypsum and halite, and irrigation return flow. The Gibbs diagram confirms the predominance of the evaporative process, while the ion exchange indices (CAI-I and CAI-II) revealed a direct Na⁺/Ca²⁺ exchange typical of areas with high agricultural activity. The majority of samples exceed the thresholds recommended by the FAO for irrigation (EC > 3000 µS/cm and often a high SAR), making rational management essential: tolerant crops, control of irrigation volumes, and regular monitoring of salinity and SAR. This high salinity limits agricultural use without proper management, restricts crop choices, and requires specific practices such as amendments, localized irrigation, or soil salinity monitoring.

According to Fig. 14, the spatial distribution of the IWQIs shows an area in the north and islands in the center, indicating lower water quality for irrigation in these sectors, hence the increased risk of soil salinization. The rest of these areas show more favorable indices. This distribution throughout the study region is graphically represented by the dataset analysis shown in Fig. 14(a), which also highlights the range of water quality values found. There is a significant variation in the water quality regulations, as seen by the IWQI readings, which range from 8 to 71 with an average of 39.5. The geospatial analysis identifies areas with water quality that may have an adverse effect on crop and soil health (Fig. 14(a)). In the region, a sizable percentage of water samples are classified as unsuitable for irrigation. This finding is significant because it raises concerns about potential negative effects on the soil’s capacity to retain moisture, distribute nutrients, and sustain crop yields overall. To counteract these consequences, eco-friendly water management strategies must be put into place. Regarding the percentage of sodium (Na%), the entire region, except for the northeastern center and the islands in the center, is characterized by high Na% values, suggesting an increased risk of sodicity, requiring amendments and modifications to agricultural practices to ensure good plant growth (Fig. 14b). Regarding soil permeability, the distribution reveals that reduced permeability mainly affects the northeast and, to a lesser extent, the center, correlated with areas of high sodicity (Fig. 14c). The distribution map of residual sodium carbonate revealed (RSC) indicated that the highest values were concentrated in the northeast and center, due to soluble carbonates in the soil at these depths (Fig. 14d). This may indicate a soil rich in limestone or marl characterized by a high carbonate content.

The sodium adsorption ratio, overall, is higher in the northeast, indicating the possible emergence of alkaline soils with reduced permeability, with a decreasing gradient towards the rest of the region (Fig. 14e). These suggest that GW in Ouled Djellal is generally suitable for irrigation. However, when saline water is used for irrigation, lower Sodium Adsorption Ratio (SAR) values are preferable. Sodium-related risks also depend on the total salt concentration in the irrigation water. Thus, water with salinity levels between 1.5 and 3.00 mS/cm and SAR > 4 should be applied cautiously to avoid soil degradation. It is important to note that soil samples must be taken annually to avoid potential salinity risks. The distribution map of soluble sodium percentage (SSP) revealed that the highest values were located in the northeast and center parts (Fig. 14f), overlapping with areas of a high SAR values (Fig. 14e), indicating problems of soluble sodium accumulation.

Fig. 13

Riverside and Wilcox Diagrams.

Fig. 14

Irrigation water quality indices (IWQIs): (a) IWQI, (b) Na%, (c) PS, (d) RSC, (e) SAR, and (f) SSP. Map created using ArcGIS Pro 2.8.8 (Esri; https://www.esri.com/arcgis/about-arcgis).

Spatial distribution maps supported with the EBKRP method

According to Fig. 15, a clear persistence of low-quality areas in the north is deduced, but with more precise boundaries and additional nuances in the quality gradient, notably unfavorable islands more distinctly identified in the center. This confirms the diagnosis of Fig. 14, while further highlighting the vulnerability of certain areas where the risk of salinization is high.

The EBKRP method allows for a much more gradual and localized decrease in SAR values across the region (Fig. 15e), which facilitates the identification of areas that are potentially affected by alkalization and reduced permeability issues due to excess sodium. The spatial distribution map of RSC provides more detailed information on the location of risks related to the presence of soluble carbonates, with maximum values in the northeast and center (Fig. 15d). The spatial distribution map of SSP defined areas of the highest proportions with soluble sodium in the northeast and center, which suggesting potential accumulation of soluble sodium in the soil, and posing risks to soil structural integrity (Fig. 15f). The comparison between Figs. 14 and 15 revealed that the EBKRP technique achieves a clear improvement through two mechanisms including refined spatial precision and optimized micro-risk zone identification. Thus, it allows more targeted management to prevent the risks of sodicity, reduced permeability, salinization, or alkalinization of soils, and therefore to adapt irrigation practices to the local reality of each sector. The superiority of the EBKRP method over the traditional IDW interpolation is quantitatively confirmed through cross-validation results (Table 7). The EBRK approach provides substantially lower error metrics (ME, RMSE, MAE) and a higher coefficient of determination (R²), reflecting better prediction reliability and accuracy.

Table 7 Cross-validation results of interpolation methods.

Fig. 15

Spatial Distribution Map of (a) IWQI, (b) Na%, (c) PS, (d) RSC, (e) SAR, and (f) SSP after using EBKRP method. Map created using ArcGIS Pro 2.8.8 (Esri; https://www.esri.com/arcgis/about-arcgis).

According to a regression-kriging approach, the EBKRP method enhances interpolation accuracy by incorporating explanatory variable rasters¹⁰⁶. Here, the spatial distribution maps derived from the EBKRP method (Fig. 15) demonstrated a significant heterogeneity in irrigation water quality across Ouled Djellal area. The IWQI indicated deterioration in water quality especially in the regions that affected by elevated levels of salinity and sodicity. Moreover, the high values of Na%, SAR, RSC, and SSP in several areas can be considered as a potential risk of soil degradation due to the accumulation of sodium, which will impact on soil structure and permeability. These findings emphasized the significance of specific water management strategies in Ouled Djellal in order to try to mitigate future hazards and to maintain viable irrigation water supplies over the long term.

IWQIs prediction using ANN, SVM and MLR

In this study, a comparative evaluation of ANN, SVM, and MLR models to predict water quality indices was realized; it revealed distinct limitations and strengths rooted in the capacity of each method to capture linear and nonlinear relationships. A number of artificial neural network (ANN) topologies designed to predict different water quality indices utilizing physicochemical factors as inputs are shown in the Fig. 16. A unique ANN model designed for a particular output index, such as the IWQI, SSP, SAR, Na, PI, and RSC, is represented by each subfigure. Measured water quality parameters including pH, EC, TDS, F⁻, Ca²⁺, Mg²⁺, HCO₃⁻, Cl⁻, Na⁺, K⁺, SO₄²⁻, and NO₃⁻ make up the input layer, which is displayed on the left with blue boxes. Green neurons, which represent one or more hidden layers and capture nonlinear interactions between the variables, process these inputs. The validation outcomes of the ANN models created to forecast several IWQIs are shown in Fig. 17. With R² values ranging from 0.94 to 0.99, all models show substantial correlations between anticipated and observed values, indicating exceptional predictive accuracy. In addition to Na% (R² = 0.95) and SSP (R² = 0.94), the WQI (R² = 0.97), SAR (R² = 0.99), PS (R² = 0.99), and RSC (R² = 0.96) models exhibit exceptionally good accuracy.

In addition, strong correlations between IWQI (R² = 0.93) and RSC (R² = 0.90) among the SVM models indicate dependable model performance (Fig. 18). While the models for SAR (R² = 0.61), SSP (R² = 0.62), and Na% (R² = 0.61) show rather poor predictive ability, PS (R² = 0.86) likewise shows respectable accuracy. Overall, even though the SVM models were able to predict the majority of indices with respectable accuracy, bigger training datasets and additional tuning could be needed to enhance the prediction performance for indices like SAR, SSP, and Na%. On the other hand, the findings of the MLR models demonstrate exceptional accuracy (R² > 0.98) for forecasting SAR, SSP, PS, RSC, and SAP, indicating that the developed model is very dependable for these parameters (Fig. 19). It exhibits moderate-to-excellent performance for sodium percentage (Na%, R² = 0.74) and good performance for the general water quality index (R² = 0.83). For further information, each model’s performance was evaluated using root mean square error (RMSE) and coefficient of determination on both training and validation datasets for IWQIs, which are compiled in Table 8.

Fig. 16

Number of Nodes for each type of activation in the ANN Model for the IWQIs indexes (SAR, IWQI, SSP, NA%, RSC, and PS).

Fig. 17

Application of the ANN model to validate output dataset for 6 indexes.

Fig. 18

Application of the SVM model to validate output dataset for 6 indexes.

Fig. 19

Application of the MLR model to validate output dataset for sex indexes.

As summarized in Table 8, the MLR approach scored almost perfect results (R² = 1.00; RMSE ≈ 0) for linearly governed indices, such as PS and RSC in the training phase, underscoring strong linear associations. Nevertheless, its predictive generalization proved limited: validation R² values declined for IWQI (R² = 0.82, RMSE = 4.56) and Na% (R² = 0.74, RMSE = 9.16), indicating an inability to accommodate the intrinsic nonlinearities inherent in certain hydrochemical interactions. For instance, in the study of¹⁰⁶ in Sidi El Hani aquifer (central-eastern Tunisia), where MLR model was satisfactory but less accurate than an ANN model, Whereas R², RMSE, and Mean Absolute Error (MAE) were 0.81, 1.20, and 1.63 for MLR Vs. 0.92, 1.02 and 0.90 for ANN. In general, MLR provided a low performed compared to other ML techniques^31,107 in hydrological and environmental modeling tasks.

Concerning the SVM model, it demonstrated an optimal balance between bias and variance. Validation (R²) values were uniformly high across indices, such as IWQI (R²= 0.93, RMSE = 2.56), SAR (R²= 0.98), PS (R²= 0.99), and SSP (R²= 0.99), which reflecting robust model generalization. However, a high value of RMSE for RSC (16.56) suggests occasional sensitivity to outliers or complex non-monotonic behavior, warranting further kernel or hyperparameter optimization to fully harness SVM nonlinear mapping capabilities. For instance⁶⁷, demonstrated that Weight of Evidence (WOE) exhibited a higher groundwater potential than SVM model according to their study in Pakistan, while¹⁰⁸ showed that XGBoost obtained a better accuracy than SVM, even if both studies stated that SVM is still a good model for water quality assessment. Conversely, our results align with the findings of¹⁰⁷ which showed that SVM model was superior than MLR and Propagation Neural Network (BPNN) for predicting of dissolved oxygen.

In our case, the ANN model outperformed in most validation scenarios, which achieving R² = 0.97 (RMSE = 1.50) for IWQI, and R² ≥ 0.99 for SAR and PS, thereby demonstrating exceptional aptitude for modeling intricate multivariate dependencies. Near-perfect training scores across all indices (R² ≈ 1.00; RMSE ≈ 0) attest to the network’s flexibility, though slight overfitting, which demonstrated by an increased validation for RSC (RMSE = 4.81), emphasizing the necessity of rigorous regularization or a higher data volume in future implementations. According to previous studies¹⁰⁹, the model performance of the ANN model outperformed like SVM, which used for quality prediction of river water in Malaysia. Nevertheless, another studies reported a contrasting result. For example¹¹⁰, reported that SVM was the most accurate model compared to ANN to predict water quality in Tireh river, IRAN. Similarly¹¹¹, indicated that the prediction result of SVM was superior to ANN model in precision and generalization ability for groundwater depth prediction in Mengcheng Country, China. Finally, the limitations of this study represented in a larger dataset may improve the resilience and precision of the machine learning predictions, even while the sample size is sufficient for the models being employed. Furthermore, the study focusses on physicochemical aspects rather than analyzing specific pollutants in depth, such as dangerous heavy metals, which might provide a more complete picture of anthropogenic stress. This study emphasizes the need for further work such as expanding the geographical area, increasing the number of samples, conducting multi-seasonal sampling campaigns, developing dynamic modelling approaches such as integrating Internet of Things (IoT) for real-time monitoring and conservation strategies.

Table 8 The outcomes of ANN, SVM and MLR models.

Conclusion

This study aimed to assess the groundwater (GW) quality of the phreatic aquifer using WQIs supported with ML models in the Ouled Djellal region, Algeria, It focused on identifying the principal geochemical mechanisms that controlled the acquisition of the GW chemistry and assessing their suitability for irrigation. This work combined chemical analyses of major ions, such as Ca²⁺, Mg²⁺, Na⁺, K⁺, SO4²⁻, HCO₃⁻, Cl⁻, and NO₃⁻ which conducted on 23 collected samples from representative boreholes. Hydrochemical characterization was complemented by chloro-alkaline indices (CAI), irrigation water quality indices (IWQs), such as IWQI, SAR, Na%, SSP, PS, RSC. Moreover, advanced multivariate statistics, such as principal component analysis and hierarchical cluster analysis were applied to discriminate sources of variability, classify water types. The results revealed dominant hydrochemical facies including calcium and magnesium chloride due to the influence of salinization and ion exchanges in the aquifer. The results of physicochemical analysis showed that the majority of the collected GW samples exhibited high mineralization, that ranging from 2566 to 16,100 µS/cm, with an average of 6994 µS/cm, due to geogenic processes (water-rock interaction and high evaporation) and the impact of human activities (intensive use of agricultural inputs). The results of IWQI revealed that about more than 60% of the samples had a restricted average quality for irrigation. The ML models, particularly the Artificial Neural Network (ANN) had the better accuracy, which showed a determination coefficient R² = 0.97 and RMSE = 1.50 in forecasting the IWQI, as well as R² values equal to or greater than 0.99 for SAR and PS, that demonstrating the robustness of the approach. Finally, integrating water quality indices and machine learning models supported with GIS techniques provides a robust, repeatable framework for effective groundwater quality forecasting, and provide a clear image for managing water quality and appropriate agriculture.