A Hybrid Approach for Assessing Aquifer Health Using the SWAT Model, Tree-Based Classification, and Deep Learning Algorithms

Abstract

Aquifer health assessment is essential for sustainable groundwater management, particularly in semi-arid regions with challenging geological conditions. This study presents a novel methodology for assessing aquifer health in the Barakar River Basin, a hard-rock terrain, by integrating tree-based classification, deep learning, and the Soil and Water Assessment Tool (SWAT) model. Employing Random Forest, Decision Tree, and Convolutional Neural Network (CNN) models, the research examines 20 influential factors, including hydrological, water quality, and socioeconomic variables, to classify aquifer health into four categories: Good, Moderately Good, Semi-Critical, and Critical. The CNN model exhibited the highest predictive accuracy, identifying 33% of the basin as having good aquifer health, while Random Forest assessed 27% as Critical heath. Pearson correlation analysis of CNN-predicted aquifer health indicates that groundwater recharge (r = 0.52), return flow (r = 0.50), and groundwater fluctuation (r = 0.48) are the most influential positive factors. Validation results showed that the CNN model performed strongly, with a precision of 0.957, Area Under the Curve–Receiver Operating Characteristic (AUC-ROC) of 0.95, and F1 score of 0.828, underscoring its reliability and robustness. Geophysical Electrical Resistivity Tomography (ERT) field surveys validated these classifications, particularly in high- and low-aquifer health zones. This study enhances understanding of aquifer dynamics and presents a robust methodology with broader applicability for sustainable groundwater management worldwide.

1. Introduction

Aquifer health assessment is a critical component in the sustainable management of water resources, particularly in regions where groundwater serves as a primary source of water for agricultural, industrial, and domestic use. The concept of Aquifer health presents an alternative paradigm, demonstrating significant potential both theoretically and practically. Aquifer health is defined here as a comprehensive measure of the stability and resilience of aquifer processes, encompassing groundwater quality, quantity, and the ecosystems dependent on them [1]. Tamblyn [1] emphasised the paradigm shift from traditional sustainability frameworks toward a more holistic and actionable health-based assessment. His work underscored the importance of evaluating aquifers not solely by their capacity to supply water but also by their ability to sustain ecological integrity and long-term human use. Building upon this foundation, the present study operationalises the concept of aquifer health through an integrated, data-driven framework that combines hydrological modelling and machine learning to classify aquifer health at a regional scale. As a recent concept, “aquifer health” is subject to scrutiny, as new ideas often invite critique. The definition of aquifer health is not universally agreed upon, and various stakeholders might have differing opinions on the health of a particular aquifer. While aquifers are typically associated with highly permeable geological formations, it is important to acknowledge that groundwater flow can also occur in less conventional settings, though the present study focuses on shallow, productive aquifers with significant yield potential [2,3]. Aquifers can be considered healthy when they maintain resilient water quality and yield productive water quantities [4,5,6]. They are characterised by indistinct boundaries and lack commercial species, making them less comparable to traditional ecosystems [7]. The exploration of commercial species typically focuses on ecosystems where biodiversity and species composition directly support economic activities, whereas aquifers function more as hydrological systems without such biological components. They are often described as hidden, enigmatic, and visually unappealing, epitomising the “black box” nature of hydrology [8]. Recharge rates are essential for maintaining aquifer levels [9]. Natural recharge processes, such as precipitation and river flow infiltration, play a significant role, while artificial methods like managed aquifer recharge can enhance groundwater replenishment [10]. Contamination from agricultural runoff, industrial discharges, and improper waste disposal introduces pollutants into groundwater, compromising water quality and posing risks to human health and ecosystems [11]. Extraction rates are another critical factor. Over-extraction for agricultural and industrial purposes leads to declining water tables and aquifer depletion, necessitating the adoption of sustainable extraction practices and policies to balance demand and supply [12]. Climate variability also impacts aquifer health, with changes in precipitation patterns and temperature affecting recharge and evaporation rates. Droughts reduce opportunities for natural recharge, while intense rainfall can increase recharge but may also lead to contamination through surface runoff [13].

Aquifers, which are underground layers of water-bearing rock, play a crucial role in maintaining the hydrological balance, and their health directly affects water availability, quality, and ecosystem stability [14]. Traditionally, aquifer health has been assessed through hydrogeological surveys, water quality testing, and the drilling of boreholes. Electrical Resistivity Tomography (ERT) has recently emerged as a valuable geophysical technique for aquifer identification, providing high-resolution imaging that reveals aquifer characteristics and delineates groundwater-bearing zones, particularly in hard-rock and heterogeneous terrains [15,16]. Despite its effectiveness, ERT is time-consuming and resource-intensive for large basins, requiring extensive field setups and data processing [17]. ERT is best suited as a complementary validation tool, alongside broader-scale models and remote sensing, to enhance aquifer health assessments in targeted areas.

In recent years, the emergence of advanced computational techniques combined with GIS has opened new pathways for advancing hydrological studies [18,19,20,21,22,23]. With its ability to handle large datasets and identify complex patterns, machine learning is a powerful tool for analysing various factors that influence aquifer health, such as precipitation, land use, and human activities. Prior scholars have used a variety of machine learning and deep learning techniques to investigate different facets of groundwater research, such as Decision Trees [24], Random Forest [25], AdaBoost [26], Support Vector Machines [27], Artificial Neural Networks [28], Gradient Boosting Machine [29], XGBoost [30], and deep learning algorithms [31,32]. Coupled with hydrological models like the Soil and Water Assessment Tool (SWAT), which is widely used for its ability to simulate detailed hydrological budget components at the watershed scale, aquifer health assessment becomes more effective [33]. SWAT provides critical insights into surface and subsurface hydrological processes, such as recharge, runoff, lateral flow, and baseflow, which are essential for understanding aquifer dynamics [13]. When integrated with machine learning, the SWAT model significantly enhances the precision of aquifer health classification, providing robust support for sustainable groundwater management in complex hydrogeological settings.

The Barakar River Basin in Jharkhand, India, presents unique challenges and opportunities for assessing aquifer health due to its diverse geology, varied land use patterns, and significant seasonal variations in rainfall. Understanding the aquifer dynamics in this context is essential for ensuring water security and supporting the socioeconomic development of the region. Numerous researchers have extensively examined groundwater conditions within the Jharkhand and Barakar River Basin [34,35,36,37,38,39,40,41,42]. These studies primarily focus on identifying groundwater potential zones, assessing vulnerability, and analysing contamination risks. A significant research gap exists regarding comprehensive methodologies for assessing aquifer health. No prior studies have investigated the integration of machine learning, deep learning, and SWAT models that incorporate hydrological, water quality, and socioeconomic variables for a comprehensive assessment of aquifer health. This gap highlights a substantial deficiency in the literature, where the multifaceted dimensions of aquifer health remain insufficiently addressed.

This research aims to bridge this gap by developing an innovative approach to aquifer health assessment in the Barakar River Basin. By integrating the SWAT model with advanced machine learning and deep learning techniques, specifically, Random Forest, Decision Tree, and Convolutional Neural Network (CNN) models, this study presents a novel framework for a comprehensive assessment of aquifer health.

2. Study Area

The Barakar River Basin, part of the more extensive Damodar River system, is situated within the Chota Nagpur Plateau and plays a crucial hydrological role in eastern India. Originating from the Hazaribagh Plateau in Jharkhand, the Barakar River flows for 256 km through a diverse landscape marked by varying topography, lithology, and geological complexity before merging with the Damodar River in West Bengal. The basin, covering an area of 6425 km², spans across Jharkhand and West Bengal, encompassing either wholly or partially seven districts in Jharkhand: Giridih, Dhanbad, Jamtara, Koderma, Chatra, Deoghar, and Hazaribagh (Figure 1). A minor section of the basin extends into the Paschim Bardhaman district of West Bengal. The elevation ranges from 97 to 1374 m above sea level, with the terrain primarily characterised by rugged topography, waterfalls, and steep slopes. The tributaries, such as the Barsoti and Usri, traverse diverse lithological formations and contribute to their complex fluvial network. The region experiences a humid subtropical monsoon climate, receiving an average annual rainfall of 1170–1383 mm, with over 80% of the rainfall occurring between June and September. The Barakar Basin’s shallow aquifers primarily develop within sandstone, alluvium, gneiss, granite, quartzite, and the banded gneissic complex (BGC), with sandstone and alluvium offering the most favourable conditions due to higher porosity and permeability, while granite and gneiss act as fractured aquifers.

Geological Setting of the Study Area

The Barakar River Basin is geologically linked to the Vindhyan and Gondwana systems of the Palaeozoic epoch, reflecting a complex and polycyclic geological history shaped by ancient cratonic stability and subsequent sedimentary and tectonic activity [43,44]. The region is predominantly underlain by the Chhotanagpur Gneissic Complex (CGC), consisting of Archean to Proterozoic granitic gneiss, granophyre, and occasional laterite capping [45]. This complex forms the structural basement and dominates the central and western portions of the basin. The Gondwana Supergroup is extensively represented, particularly in the southeastern parts of the basin, by the Kharharbari, Barakar, Barren Measures, and Raniganj Formations [46,47]. Among them, the Barakar Formation is especially notable for its coarse-grained sandstone, coal seams, and fluvio-deltaic depositional signatures, which provide critical aquifer zones [47]. These formations originated in fault-controlled half-graben basins during the Late Carboniferous to Early Permian, indicative of crustal extension and block faulting associated with Gondwana rifting. Volcanic and metavolcanic units such as the Dhanjori and Dalma Lava Formations are present primarily in downstream areas. These Proterozoic basic and ultrabasic rocks, along with basic lava flows and intrusive rocks in the eastern and northwestern parts of the basin, are interpreted as remnants of ancient mafic magmatism and rift-related volcanic episodes [48]. Mica schists, unclassified metamorphic rocks, and quartzites are scattered across the terrain, reflecting medium to high-grade metamorphism. The Bihar Mica Belt, rich in muscovite-bearing phyllite, schist, and amphibolite, is a prominent litho-tectonic feature associated with late Proterozoic pegmatitic intrusions and post-collisional tectonics [49]. These tectono-sedimentary events have played a fundamental role in shaping the basin’s groundwater dynamics and geological framework, as further explored in the lithology and hydrogeological sections of this study.

3. Materials and Methods

3.1. Data Inputs and Thematic Map Preparation

This study integrates a wide array of topographic, geospatial, and environmental information to train machine learning and deep learning models for accurate mapping of aquifer health. A total of 20 thematic datasets were employed, comprising 14 hydrological variables, a water quality index, and five socioeconomic factors. The slope data for topographical analysis were derived from a 30 m resolution Shuttle Radar Topography Mission (SRTM) Digital Elevation (DEM) using ArcGIS 10.5. Lithological and lineament density data were obtained from the Geological Survey of India’s Bhukosh portal (https://bhukosh.gsi.gov.in/Bhukosh/Public; accessed on 12 March 2024). Groundwater data covering extraction for various uses (domestic, irrigation, industrial) and the stage of groundwater extraction were sourced from the Central Ground Water Board [50], which includes block-level data across 36 blocks partially or entirely within the basin. A soil texture map at a 1:50,000 scale was obtained from the National Bureau of Soil Survey and Land Use Planning (NBSS-LUP). The spatial distribution of pond locations was derived from Survey of India (SOI) Toposheets and Google Earth imagery, with pond midpoints used to generate a density map in ArcGIS [13]. This approach considers the number of ponds per km² but does not account for individual pond size or volume due to data limitations. The influence of pond size and volume on groundwater recharge capacity is acknowledged, and it is recommended that this aspect be explored in future studies to improve the accuracy of recharge assessments. The Land use map was prepared using Esri Sentinel-2 data. A complete summary of data sources is presented in Table S1 (see Supplementary File). The administrative boundary shapefiles for India and Jharkhand used in the location map were obtained from the Survey of India’s official website (https://onlinemaps.surveyofindia.gov.in). The classification of aquifer media in the study area was based on the integration of geological and hydrogeological maps available through the India Water Resources Information System (https://indiawris.gov.in/wris/#/, accessed on 3 April 2025), developed by the Central Ground Water Board (CGWB). Aquifer depth was not considered in this study due to the unavailability of detailed depth data.

In this study, dug wells were considered, which are generally constructed at shallow depths and are known to tap the near-surface aquifer. The groundwater level and quality assessment and associated modelling were conducted specifically for the shallow aquifer system. The selection of dug wells was based on their widespread use for domestic and agricultural purposes in the study area, as well as the availability and consistency of water quality data from these sources. Although it is acknowledged that the region may host multiple aquifer layers with varying geochemical characteristics, a stratified analysis was not performed due to limitations in well construction data (screen depth, aquifer depth, lithological logs), which were not uniformly available for all sampling locations. The focus on the shallow aquifer is justified by its greater vulnerability to surface-derived contamination and its critical role in sustaining rural water supply [51]. During the pre-monsoon and post-monsoon periods of 2023, an extensive survey was conducted to measure groundwater levels in 200 dug wells across the study area. To ensure broad and representative coverage, the survey used a combination of random sampling complemented by stratified sampling to address geographic and hydrologic variability across the region. Changes in groundwater levels were then assessed by comparing the data from the pre-monsoonal and post-monsoonal measurements.

All groundwater samples used in this study were obtained from dug wells tapping the shallow aquifer system, ensuring consistency in the aquifer type for quality and productivity assessments. Physical parameters such as TDS, EC, and pH were measured directly in the field for water quality assessment using a Hanna multiparameter instrument (HI98194; Hanna Instruments, Woonsocket, RI, USA). The total alkalinity of each sample was determined on-site with an AQUASOL Alkalinity Kit (AE-214; AQUASOL, Mumbai, Maharashtra, India), while fluoride levels were assessed using an AQUASOL Fluoride Kit (AE-210; AQUASOL, Mumbai, Maharashtra, India). Water samples were collected from the dug wells and stored in 500 mL bottles for further laboratory examination. In the lab, various instruments were used to measure cationic and anionic concentrations, including Mg²⁺, Na⁺, K⁺, Ca²⁺, HCO₃⁻, Cl⁻, SO₄²⁻, F⁻, and NO₃⁻, as well as pH, electrical conductivity (EC), and total dissolved solids (TDS). A digital flame photometer was used to test the concentrations of Na⁺ and K⁺, the argentometric method was used to detect the levels of Cl⁻, and titrimetric techniques were used to examine the Ca²⁺ and Mg²⁺ concentrations.

Table 1. Descriptive statistics of groundwater quality parameters.

Water Quality Parameters	Minimum	Maximum	Mean	Median	Standard Deviation	Kurtosis	Skewness
EC (μS/cm)	306.25	1217.42	657.24	619.94	164.21	0.95	0.76
TDS (mg/L)	199.06	791.32	427.07	402.97	106.70	0.96	0.77
pH	7.46	8.60	8.14	8.15	0.15	2.62	−0.53
HCO3− (mg/L)	93.29	335.39	185.12	186.11	39.96	1.27	0.59
Cl− (mg/L)	15.96	207.91	80.86	79.66	28.96	3.59	1.18
F− (mg/L)	0.07	1.33	0.61	0.58	0.17	2.42	0.79
NO3− (mg/L)	0.00	160.58	35.24	28.61	25.57	3.07	1.40
SO42− (mg/L)	3.68	85.85	36.28	36.20	20.41	−1.20	0.26
TH (mg/L)	100.12	422.31	222.35	214.41	54.22	1.84	1.06
Ca2+ (mg/L)	16.04	95.86	47.59	46.16	10.49	3.51	0.94
Mg2+ (mg/L)	4.96	64.95	25.12	23.82	10.43	0.89	0.86
Na+ (mg/L)	9.23	134.29	50.31	44.61	20.53	0.95	0.88
K+ (mg/L)	0.47	17.01	5.00	4.92	2.62	5.52	1.74

presents the descriptive statistics for the different cationic, anionic, and physical parameters for the study area. The geochemical data were plotted on a Piper diagram [52], providing insights into the groundwater facies. The results were compared with the WHO [53] recommended standards for drinking water quality. The charge balance error (CBE) was calculated using the following equation:

$$\%\mathrm{C}\mathrm{B}\mathrm{E}={\displaystyle \frac{\sum Cations-\sum Anions}{\sum Cations+\sum Anions}}\times 100$$

(1)

In this study, all cationic and anionic concentrations are presented in mg/l. The CBE for all the samples collected was found to fall within the acceptable range, specifically, within ±10%.

The Water Quality Index (WQI) has been assessed according to Equation (2) [54,55,56,57].

$$\mathrm{W}\mathrm{Q}\mathrm{I}=\sum {SI}_i$$

(2)

where ${SI}_i$ is the sub-index value for the i-th variable, calculated as the product of the relative weight $W_i$ and the water quality rating $Q_i$’

$${SI}_i=W_i\times Q_i$$

(3)

Relative weight ($W_i$) has been calculated using the following Equation (4).

$$W_i=w_i/{\sum }_{i=1}^n{w}_i$$

(4)

where $w_i$ is the initial weight assigned to the i-th parameter based on its relative importance to water quality (see Supplementary File, Table S2).

$$Q_i=(C_i/S_i)\times 100$$

(5)

where $Q_i$ represents the ratio of the measured parameter concentration to the standard limit for the corresponding parameter (Equation (5)). $C_i$ is the measured concentration, and $S_i$ is the standard limit for drinking water quality.

For model training and testing, data from 898 wells were obtained from the Jharkhand Geospatial Portal, managed by the Jharkhand Space Applications Center (https://gis1.jharkhand.gov.in/Geojharkhand/?mbstring=2952c22180808d8457ca04307de267bd, accessed on 3 April 2025). According to previous studies, wells yielding more than 50 litres per min (LPM) are considered productive and are indicative of favourable aquifer conditions [58,59,60]. This threshold was used to assess well productivity and the health of the aquifer. The data were divided into training and testing datasets using a 70:30 ratio to balance model accuracy with generalisability, ensuring robust performance in yield prediction and aquifer assessment (Figure 2). Diverse sampling techniques guided the choice during dataset preparation, which included several iterations to create three separate training and testing datasets. The training and testing datasets were constructed with specific conditions to ensure representativeness: 70% of the data points for training were selected from the productive (high-yield) and poor-yield categories, while 30% from each category was allocated for testing. This approach was designed to minimise potential biases and ensure that the results were robust and not influenced by random variation [61].

The probabilistic results produced by the system will be used to create maps that highlight areas with different levels of aquifer health, ranging from good to critical. The Jenks natural breaks method was used to divide the aquifer health map into four distinct categories [62]. Figure 3 presents the flowchart outlining the methodology used in the study. Each map was meticulously generated at a 30 m spatial resolution and accurately projected using the WGS 1984 UTM Zone 45N coordinate system, ensuring consistency and geospatial precision across the dataset.

3.2. Methods for Electrical Resistivity Tomography (ERT)

ERT is a widely recognised geoelectrical technique used to investigate subsurface structures by analysing resistivity contrasts. Traditional surveys utilise electrode configurations to produce resistivity curves, profiles, or pseudo-sections, providing insights into vertical and horizontal variations [63,64]. Based on Ohm’s law, ERT assesses subsurface properties influenced by porosity, permeability, rock composition, and water content [65]. This study adopts the dipole–dipole array, a configuration that enhances subsurface imaging for aquifer analysis. The method utilises two pairs of electrodes: one pair injects electrical current into the ground, while the other measures the resultant voltage [66]. The dipole–dipole configuration is especially effective in detecting lateral resistivity variations, making it highly suitable for mapping intricate geological formations, locating groundwater resources, identifying contamination plumes, and evaluating subsurface hazards [67,68]. Four locations were selected for ERT surveys based on the CNN-generated aquifer health assessment map—two in zones with good aquifer health and two in areas identified as having critical aquifer health risks. This selection was intended to capture contrasting subsurface conditions and validate the spatial variability of aquifer characteristics inferred from the model.

3.3. Methods for SWAT Modelling, Tree-Based Classification, and Deep Learning Approaches

3.3.1. Soil and Water Assessment Tool

The Soil and Water Assessment Tool (SWAT) is a semi-distributed, physically based hydrological model developed by the United States Department of Agriculture (USDA) to predict the impact of land management practices on water, sediment, and agricultural chemical yields in complex watersheds with varying soils, land use, and management conditions over long periods [69]. SWAT operates on a daily time step and is particularly well-suited for analysing hydrological processes in a river basin. The model divides a watershed into multiple sub-basins, which are further partitioned into Hydrologic Response Units (HRUs). Each HRU represents a unique combination of land use, soil type, and slope class, allowing for detailed spatial representation of watershed characteristics [70]. This approach enables SWAT to simulate the heterogeneity of hydrological processes across the landscape without requiring full spatial discretisation.

The SWAT model was applied to assess the hydrological budget components of the Barakar River Basin from 2003 to 2022, with a focus on parameters relevant to aquifer health. The watershed was divided into 32 sub-watersheds and 236 HRUs. A 30 m SRTM DEM provided topographic data, classifying slopes into gentle (0–3%), moderate (3–15%), and steep (>15%) gradients. Soil information was sourced from the FAO-UN Harmonized World Soil Database. Daily meteorological data for the period 1984–2022, including precipitation, temperature, solar radiation, relative humidity, and wind speed, were obtained from the NASA Prediction of Worldwide Energy Resources (POWER) database. This dataset is globally recognised for its reliability and suitability in hydrological and environmental modelling, particularly in regions with sparse ground-based observations. The NASA POWER data are derived from satellite observations and assimilated reanalysis products, and are provided at a spatial resolution of 0.5° × 0.625°. While this resolution is appropriate for large-scale watershed studies, it may not fully capture microclimatic variations in heterogeneous or topographically complex terrains. Despite this limitation, previous studies have demonstrated its robustness for long-term trend analysis and water resource assessments [71,72,73]. The use of this dataset is justified given its temporal continuity, comprehensive coverage, and consistency over the multi-decadal study period [74]. The SWAT model generated hydrological outputs essential for aquifer assessment, including lateral flow, groundwater recharge, surface runoff, base flow, return flow, and soil water content. The distribution of these variables by sub-watershed is provided in the figures and text, detailing how these variables vary spatially across different regions of the basin. All hydro-environmental input parameters used in the SWAT model were computed as annual mean values using a 20-year simulation period (2003–2022). This approach accounts for interannual variability and ensures a more stable representation of long-term hydrological conditions across the sub-basins. The detailed methodology and equations used for these calculations are available in our previous work [13]. This SWAT-based approach supports the integration of hydrological insights with machine learning and deep learning models to classify aquifer health across the basin.

3.3.2. Decision Tree

This study employs the Decision Tree algorithm as a core machine learning methodology for its effectiveness in handling complex groundwater data and providing interpretable models. A Decision Tree functions by iteratively dividing the dataset into branches according to feature values, forming a tree structure. Each branch corresponds to a specific decision rule, while each leaf node signifies an outcome or classification [75]. The maximum depth for hyperparameter tuning in this study has been set to 3. This structure enables the model to handle both numerical and categorical data effectively, making it highly suitable for aquifer health assessment, particularly in cases where multiple conditioning factors are present. By determining the best splits using criteria such as Gini impurity or information gain, the model effectively isolated the most influential factors for each branch [76]. The Decision Tree’s interpretability also provided insights into how each factor contributed to the classification of aquifer health.

3.3.3. Random Forest

The Random Forest (RF) algorithm is a robust tool widely used in groundwater studies [77,78]. It employs a bagging method, randomly dividing the training data to create multiple independent decision trees that optimise classification. Data points not included in model training are placed into an “out-of-bag” (OB) set, which helps assess the model’s accuracy using these excluded samples [79,80]. Two key parameters were employed for optimising the model: the number of trees (nTree) and the number of randomly selected variables (mTry). These play a vital role in enhancing the accuracy of the Random Forest (RF) model. In this study, a range of values was explored, with nTree tested between 100 and 1000 and mTry varying from 1 to 10. The optimal configuration was determined to be nTree set at 500 and mTry at 5, substantially improving the model’s performance. Adjusting these parameters proved instrumental in increasing the model’s precision and effectiveness in data classification.

3.3.4. Convolutional Neural Network (CNN)

Convolutional Neural Network (CNN) is applied as a deep learning methodology for analysing spatial patterns and predicting aquifer health across the study area. CNN is structured with convolutional layers that use filters to scan the input data, pooling layers to decrease dimensionality, and fully connected layers that analyse the extracted features to perform classification or regression tasks [81]. This structure enables CNN to learn complex spatial dependencies within the data, making them particularly suitable for geospatial and hydrological pattern tasks. The CNN model in this study was trained on various spatial conditioning factors, which were represented as spatial grids or maps [82]. By iteratively updating its filters through backpropagation, a CNN can learn and recognise patterns that correlate with healthier or more vulnerable aquifer areas [83]. This deep learning approach enhances prediction accuracy by leveraging the spatial relationships among variables, providing a robust model for groundwater management applications. For hyperparameter testing in this research, a 1D CNN was implemented. The filter size was set to 128, with a kernel size of 3, and the ReLU function was chosen as the activation function. A pooling size of 2 was selected for the 1D max-pooling layer.

3.4. Multicollinearity of Aquifer Health Conditioning Factors

A multicollinearity analysis was performed on the 20 variables selected for aquifer health assessment in the Barakar River Basin to ensure their independence and mitigate the risk of model instability. Multicollinearity occurs when predictor variables are highly correlated, leading to unreliable estimates and potentially misleading interpretations. Multicollinearity is indicated when the tolerance value is less than 0.10 and the Variance Inflation Factor (VIF) exceeds 10 [84,85]. This work computed the tolerance value and VIF from the following equations to assess multicollinearity:

Tolerance = (1 − Ri²)

(6)

VIF = 1/(1 − Ri²)

(7)

where Ri represents the regression coefficient.

In this study, all Tolerance values were above 0.38, with the most significantly higher values, indicating no multicollinearity issues (Table S3, see Supplementary File). VIF values remained well below the critical threshold of 10, with none exceeding 4, further supporting the absence of multicollinearity. The highest VIF values were found for precipitation (3.81) and groundwater extraction for domestic use (3.82), both of which are well within acceptable limits, indicating only mild collinearity. These results confirm that each variable provides unique and independent information to the model, ensuring stable and reliable estimates without inflated standard errors or bias. The variables used in this study are free from redundancy and multicollinearity, preserving the model’s interpretability.

3.5. Rationale for the Hybrid Modelling Framework

To address the spatial and temporal complexities of aquifer health in the Barakar River Basin, a hybrid methodology combining the SWAT model, tree-based classifiers, and deep learning was adopted. While field-based methods remain essential for local-scale hydrogeological investigations, the integrated approach was selected to overcome challenges posed by complex spatial heterogeneity, limited monitoring infrastructure, and the multivariate nature of factors influencing aquifer health across a large basin area [86,87]. The availability of long-term hydrometeorological and geospatial datasets enabled the simulation of key groundwater processes affecting aquifer conditions. The application of data-driven models such as Random Forest and CNN facilitated the integration of multiple nonlinear variables, including hydrological, geochemical, and anthropogenic parameters, thus supporting the robust prediction and classification of aquifer health. This hybrid framework allows for scalable, reproducible assessments that synthesise hydrological dynamics, water quality indicators, and socioeconomic driver factors that are often difficult to evaluate holistically using field data alone. Various machine learning models were used to assess aquifer health, aiming to compare the model performance and sensitivity to hydrogeological factors. While the differences between models are minor, they remain relevant, reflecting unique interpretations that contribute to more accurate and reliable predictions.

4. Results and Discussion

4.1. Hydrogeological Conditioning Variables

4.1.1. Principal Aquifer Media

The importance of the principal aquifer media in assessing aquifer health stems from their critical role in determining groundwater storage, movement, and quality. In the Barakar Basin, six key aquifer media have been identified: alluvium, quartzite, gneiss, the banded gneissic complex, sandstone, and granite (Figure 4a). Sandstone, known for its high porosity and permeability, serves as a good aquifer with significant recharge potential [88]. Alluvium, composed of unconsolidated sediments like sand and gravel, exhibits high porosity and permeability, allowing for efficient water absorption and transmission. Low porosity and permeability in media such as granite, certain types of gneiss, and quartzite can restrict these processes, making groundwater storage more dependent on secondary porosity from fractures and weathering [89]. The banded gneissic complex (BGC) exhibits variable permeability, influenced by the degree of fracturing and foliation, with the recharge potential ranging from limited to moderate [90]. A significant portion of the basin consists of BGC (78%) and gneiss (16%), while in the upper catchment, alluvium, quartzite, and granite are found in some small pockets.

4.1.2. Lineament Density

Lineament density (LD) determines the subsurface’s capacity to facilitate water infiltration and groundwater recharge. Lineaments, which include fractures, faults, and joints, serve as pathways for the movement of water. Areas with higher LD typically exhibit better subsurface connectivity, enhancing groundwater flow and recharge, which is essential for maintaining a sustainable groundwater supply, particularly in regions reliant on aquifers [91]. Areas with low LD may experience poor recharge, resulting in diminished aquifer health, reduced water availability, and increased vulnerability to groundwater depletion. Within the basin, LD ranges from 0 to 0.94 km/km² (Figure 4b). The majority of the area (61%) is characterised by low LD (0–0.07 km/km²), while high LD (0.37–0.94 km/km²) is observed in 5% of the basin, and moderate LD (0.20–0.37 km/km²) is found in 11% of the area.

4.1.3. Soil Texture

The size and composition of soil particles play a crucial role in determining soil porosity and permeability, which in turn directly influence groundwater recharge rates. Fine-textured soils, such as fine silty and fine loamy soils, consist of smaller particles that compact tightly, limiting pore space and reducing the ease with which water can percolate through, thereby hindering recharge. Coarse-textured soils, like sandy loam, loamy skeletal, and coarse loamy, have larger particles and greater pore spaces, allowing for more rapid infiltration and more efficient aquifer recharge [92]. In this basin, fine silty and fine loamy soils are the most prevalent, covering 43% and 28% of the area, respectively. Sandy loam and loamy soils are primarily concentrated in the lower catchment areas. In contrast, coarse loamy soils are predominantly found in the upper and middle catchment regions (Figure 4c). Loamy skeletal soils account for 13% of the area.

4.1.4. Slope

Topographical slope influences surface runoff, infiltration rates, and groundwater recharge. Areas with gentle slopes typically allow more water to infiltrate into the ground, enhancing the recharge potential and supporting aquifer health. Steep slopes can boost surface runoff, therefore limiting the water available for groundwater replenishment [93]. In the basin, a significant portion (65%) is characterised by gentle slopes ranging from 0° to 3°, while 23% of the area has moderate slopes between 3° and 5° (Figure 4d). A tiny portion (3%) of the basin features steep slopes exceeding 10°.

4.1.5. Lithology

Lithology, which examines the physical characteristics of rocks and sediments, is essential for aquifer health analysis, as it determines groundwater storage, movement, and recharge potential. A total of 10 distinct lithological units are present in this basin (Figure 4e). Older alluvium, mainly concentrated in the upper catchment, is characterised by high porosity and permeability, making it highly conducive to groundwater recharge. Schists can facilitate water percolation, especially when foliated and fractured, offering relatively good recharge potential [94]. The Bihar Mica Belt, comprising phyllite, schist, and amphibolite, generally has moderate recharge potential due to foliation and fractures in the schists and phyllites. The Lower Gondwana Group includes both permeable sandstones and less permeable shales [95]. The presence of sandstone enhances recharge, while shale acts as a barrier. This mixed lithology provides moderate to high recharge potential. The Chhotanagpur Gneissic Complex, which covers a substantial portion of the basin (79%), comprises various types of gneiss that exhibit moderate permeability, depending on weathering and fracturing, resulting in a moderate recharge potential [96]. Unclassified metamorphics encompass a diverse range of rock types, some of which may exhibit moderate permeability when fractured, allowing for moderate recharge potential [97]. Basic Lava (mainly basalt), concentrated in the lower catchment, exhibits secondary porosity when weathered or fractured, allowing for moderate groundwater recharge. Pegmatite and quartz veins, however, are characterised by low permeability and porosity, which restricts groundwater recharge. Both soil texture and lithology impact groundwater; soil texture primarily affects surface and shallow aquifer recharge, while lithology governs deeper aquifer behaviour.

4.1.6. Precipitation

The amount and distribution of precipitation directly determine the volume of water that can be stored within an aquifer. Consistent and sufficient rainfall helps maintain aquifer recharge, supporting the sustainability of water supplies, particularly in areas where groundwater is the key source. Conversely, areas with low precipitation are more susceptible to groundwater depletion, leading to poor aquifer health and potential water scarcity. Annual precipitation in this basin ranges from 1170 to 1384 mm (Figure 4f). The relatively higher precipitation (1242–1384 mm) is concentrated in the lowermost catchment, covering approximately 15% of the area. Precipitation gradually decreases towards the western part of the basin, with the uppermost catchment receiving the lowest rainfall, ranging from 1170 to 1185 mm.

4.1.7. Surface Runoff

Effective aquifer health management necessitates balancing surface runoff and infiltration rates to sustain adequate groundwater recharge. Surface runoff levels significantly impact aquifer recharge, as high runoff reduces the amount of water available to percolate into the subsurface and replenish groundwater reserves [13]. Surface runoff in this basin ranges from 246 to 724 mm (Figure 4g). High to moderately high runoff (429–724 mm) is concentrated in the middle and lower catchments, covering approximately 46% of the area. Low to moderately low runoff (246–373 mm) is predominantly located in the upper and middle portions of the basin. Low runoff conditions enhance recharge potential, supporting healthier aquifer systems.

4.1.8. Groundwater Recharge

Groundwater recharge is crucial for maintaining the balance of groundwater resources, ensuring the aquifer remains sustainable over time. Adequate recharge is crucial for preventing groundwater depletion, maintaining water quality, and sustaining ecosystems that rely on stable groundwater levels. Higher recharge rates help replenish the aquifer and sustainably ensure water availability for various needs. Lower recharge rates can result in declining groundwater levels, increased vulnerability to drought, and poor aquifer health [98]. The annual groundwater recharge in this basin ranges from 52 to 413 mm (Figure 4h). Higher recharge (323–413 mm) is concentrated in the upper and middle catchments, covering 21% of the study area. Low annual groundwater recharge (52–65 mm) is also focused in the upper catchment, encompassing 12% of the area.

4.1.9. Soil Water Content

Soil water content refers to the volume of water retained in the soil after precipitation, influencing the availability of water for groundwater recharge. The soil’s capacity to retain and release water directly impacts the amount of water that percolates through to recharge aquifers. Higher soil water content generally promotes groundwater recharge, supporting healthier aquifer conditions [99]. Low soil water content can limit recharge potential, leading to a decline in aquifer health over time. In this basin, soil water content ranges between 87 and 258 mm (Figure 4i). The highest soil water content (216–258 mm) is concentrated in the middle section of the basin, covering 20% of the study area, while the lowest soil water content (87–110 mm) is found in the basin’s uppermost region.

4.1.10. Lateral Flow

Subsurface lateral flow is a dynamic component of groundwater recharge, especially in areas where vertical infiltration is less effective. Higher lateral flow allows more water to reach the aquifer, enhancing recharge potential and supporting healthier aquifer conditions [100]. Low lateral flow restricts the amount of water available for recharge, potentially leading to declining aquifer health. Lateral flow is highest (33–46 mm) in the southern part of the lower catchment, covering 8% of the area (Figure 4j). Lateral flow is lowest (1–3 mm) in the northern part of the basin, encompassing 37% of the study area. Moderately high lateral flow (13–33 mm) is observed in the upper catchments of the basin, covering 21% of the study area.

4.1.11. Base Flow

Maintaining aquifer health, especially during dry periods, relies heavily on subsurface base flow. A high base flow guarantees a steady supply of groundwater to both aquifers and surface water bodies, thereby supporting long-term sustainability [101]. Low base flow can lead to declining groundwater levels and negatively impact the health of aquifers. In this basin, base flow ranges from 28 to 370 mm (Figure 4k). A significant amount of base flow (285–370 mm) is predominantly found in the upper catchment area, accounting for 21% of the region. A lower base flow (28–40 mm) is also present in the upper catchment, comprising 12% of the study area. Moderately high base flow (200–285 mm) is concentrated in both the lower and upper catchments of the basin.

4.1.12. Return Flow

Return flow from the shallow aquifer refers to the movement of water from shallow groundwater sources back into the surface water. This process helps maintain soil moisture, reduces erosion, and enhances water quality through natural filtration. High return flow increases aquifer recharge potential, providing a buffer against drought and sustaining local ecosystems reliant on consistent groundwater supplies. Low return flow can lead to declining groundwater levels and increased vulnerability to drought. Return flow is relatively higher in the upper catchments of the basin, gradually decreasing towards the lower catchments (Figure 4l).

4.1.13. Groundwater Level Fluctuation

Groundwater level fluctuation can signal instability within an aquifer system. Such instability often arises from excessive groundwater extraction, insufficient recharge, or changes in land use practices, leading to stress on the aquifer. This stress can compromise the aquifer’s ability to maintain consistent water availability, potentially impacting both ecological systems and human use. Low groundwater level fluctuations are generally associated with a more stable aquifer system [102]. This stability suggests that the aquifer maintains a balanced relationship between recharge and discharge, indicating sufficient groundwater storage throughout the year. A stable aquifer is better equipped to handle periods of drought and can more effectively support the ecosystems and human activities that rely on its resources. The fluctuation ranges in this basin are between 1 and 6.5 m (Figure 4m). High fluctuations, ranging from 4 to 6.5 m, are predominantly found in the central portion of the basin, accounting for 34% of the study area. Low fluctuations, measuring between 1 and 3 m, are primarily concentrated in the upper catchment part, which comprises 32% of the study area.

4.1.14. Pond Density

More water could permeate into the soil and replenish aquifers when the pond density is high, which helps improve groundwater recharge. This replenishment is crucial for maintaining sustainable groundwater levels and supporting ecosystems reliant on aquifers. Increased pond density can enhance local microclimates and promote biodiversity, contributing to the overall resilience of the ecosystem. Low pond density can lead to decreased groundwater recharge, potentially resulting in lower aquifer levels and compromised water quality [13]. The pond density in this basin ranges from 0 to 10 ponds/km² (Figure 4n). The majority of the study area exhibits low pond density, with values ranging from 0 to 1 pond/km². High pond density, defined as 3–10 ponds/km², is only scattered across 1% of the basin. Moderate density (1–3 ponds/km²) is concentrated in 16% of the study area.

4.2. Water Quality Parameters

Water Quality Index

Assessing aquifer health is crucial for sustainable water management, especially in areas where groundwater serves as a primary water source. A range of water quality parameters helps to determine the chemical and physical properties of groundwater, offering insights into potential contamination and the overall usability of the water. Among these parameters, EC and TDS are key indicators. EC, with an average value of 657 µS/cm in the study area, reflects the concentration of dissolved ions, which may indicate contamination from mineral dissolution or human activities. TDS averaged 427 mg/L, providing a general measure of the water’s dissolved substances, which affect both taste and usability, especially for drinking purposes. The pH of groundwater averaged around 8, indicating slightly alkaline conditions, which is within the safe range (6.5–8.5) for most uses. Deviations from this range could affect water quality by altering the chemical balance. TH, averaging 222 mg/L, is another essential factor, as it influences the scaling potential of water in pipes and appliances, potentially impacting household and industrial use. High hardness levels can lead to scaling, reducing the efficiency of water systems. Other major ions, such as HCO₃⁻, Cl⁻, Ca²⁺, Mg²⁺, Na⁺, K⁺, SO₄²⁻, and NO₃⁻, also play a significant role in determining aquifer health. Bicarbonate levels, averaging 185 mg/L, contribute to the alkalinity of water, while Cl⁻, with a mean concentration of 81 mg/L, can signal pollution from agricultural or industrial sources if elevated. Ca²⁺ and Mg²⁺, averaging 48 mg/L and 25 mg/L, respectively, are major contributors to water hardness. Na⁺ and K⁺, with mean values of 50 mg/L and 5 mg/L, respectively, may reflect agricultural runoff. SO₄²⁻ and NO₃⁻ levels, averaging 36 mg/L and 35 mg/L, respectively, are important for understanding contamination, especially nitrate, which is often linked to agricultural activities and wastewater. Fluoride, with a mean concentration of 1 mg/L, is within acceptable limits but needs monitoring, as excessive fluoride can pose health risks. Lithology, hydrogeological conditions, anthropogenic activities, and climate influence water quality. The dominant water facies is likely calcium–magnesium–bicarbonate, influenced by carbonate-rich rocks (Figure 5). However, sodium–chloride facies may occur in areas with salt-rich formations or in the vicinity of anthropogenic activities. The water type can vary from calcium–bicarbonate to sodium–chloride, depending on the dominant ions.

Several key water quality parameters, including Mg²⁺, Na⁺, K⁺, Ca²⁺, HCO₃⁻, Cl⁻, SO₄²⁻, F⁻, NO₃⁻, pH, EC, TDS, and TH, are essential for calculating the Water Quality Index (WQI). These parameters provide valuable insights into the chemical composition of groundwater, and their levels are closely linked to aquifer health on a basin-wide scale. The WQI aggregates complex water quality data into a single score, allowing stakeholders to interpret the overall condition of an aquifer easily. By applying the WQI, it becomes possible to identify specific contaminants and their concentrations, which is crucial for pinpointing pollution sources and implementing effective mitigation strategies. Regular monitoring of the WQI also facilitates tracking of water quality changes over time, a necessary measure for understanding the impacts of land use changes, agriculture, industrial activity, and climate variability on aquifer health. High WQI scores signify safe water for consumption, while low scores may indicate potential health risks. The WQI is categorised into five classifications: excellent (0–25), good (25–50), poor (50–75), very poor (75–100), and unsuitable for drinking (>100). In the study area, the calculated WQI values ranged from 34 to 134 (Figure 6). The data reveal that none of the water samples fall into the excellent category (Table S4, see Supplementary File). Only 17.5% of sample locations achieved a good WQI, primarily located in the northern part of the basin. A majority of locations, 59%, fall under the poor category, while 19.5% are classified as very poor. Eight sample locations, concentrated in the Koderma, Pirtanr, and Deori blocks, showed WQI scores above 100, indicating unsuitability for drinking. However, this water may still be viable for irrigation and specific domestic uses.

4.3. Socioeconomic Variables

4.3.1. Groundwater Extraction for Domestic Purposes

Groundwater extraction for domestic purposes refers to the volume withdrawn for residential use, which can significantly impact aquifer levels and recharge potential. Excessive extraction can lead to water quality depletion and long-term sustainability challenges, whereas sustainable practices contribute to maintaining aquifer health and supporting recharge processes. High extraction rates (9041–24,137 m³/day) are observed in the lower and middle catchments of the basin, encompassing 31% of the area, particularly within the Giridih, Nirsha, Dhanwar, Jamua, Dumri, Jamtara, and Salanpur blocks (Figure 7a). Low extraction rates (2384–3699 m³/day) are predominantly found in the upper catchments, notably, in the Mayurhand, Padma, Chalkusa, Tati Jharia, and Itkhori blocks.

4.3.2. Groundwater Extraction for Irrigation Purposes

The extent and rate of groundwater extraction for irrigation have a direct impact on aquifer health, recharge capacity, and the long-term availability of water resources. When managed sustainably, groundwater extraction can meet crop water needs without overexerting the aquifer system. Over-extraction for irrigation can lead to rapid declines in water tables and reduced recharge rates, particularly in areas with limited natural replenishment [103]. High extraction rates for irrigation, ranging from 24,384 to 35,233 m³/day, are observed in the northern–central catchments of the basin, which cover 12% of the area, particularly in the Dhanwar and Jamua blocks (Figure 7b). The largest portion of the basin (35%) experiences extraction rates of 15,726–24,384 m³/day, while lower extraction rates are concentrated in 18% of the basin area, mainly in the Padma, Koderma, Bagodar, Chalkusa, Purbi Tundi, Jamtara, Domchanch, and Markacho blocks.

4.3.3. Groundwater Extraction for Industrial Purposes

Industrial activities often require significant volumes of groundwater, which can substantially impact aquifer levels, recharge potential, and overall sustainability. The scale of groundwater extraction for industrial use plays a critical role in determining aquifer health, with higher extraction levels posing increased risks of depletion and contamination [104]. The majority of the basin (72%) experiences low groundwater extraction rates of less than 274 m³/day for industrial purposes (Figure 7c). Higher extraction rates, exceeding 2740 m³/day, are concentrated in a few blocks within the basin, particularly in Giridih, Gobindpur, Nirsha, and Topchanchi, covering approximately 8% of the area.

4.3.4. Stage of Groundwater Extraction

The stage of groundwater extraction indicates the proportion of available groundwater resources currently in use. A higher stage of extraction suggests that groundwater is being withdrawn at a rate potentially exceeding natural recharge, leading to declining water tables, reduced water availability, and possible aquifer degradation. The critical stage of groundwater extraction (>90%) is concentrated primarily in the Jainagar and Topchanchi blocks (Figure 7d). The semi-critical stage (70–90%) is observed in the Gobindpur, Giridih, and Koderma blocks. Groundwater extraction remains at a very safe stage (18–30%) across 25% of the basin area, particularly in the Sariya, Bagodar, Pirtanr, Chauparan, Padma, Gande, and Pubi Tundi blocks.

4.3.5. Land Use and Land Cover (LULC)

The influence of land cover types on water infiltration, evapotranspiration, and aquifer recharge capacity is significant, as each category affects groundwater sustainability in distinct ways. Within this basin, six primary LULC classes are identified, each contributing variably to aquifer health (Figure 7e). Water bodies, including lakes, ponds, and rivers, are direct contributors to groundwater recharge, as percolation from these sources effectively replenishes aquifers [9]. Forested areas typically enhance groundwater recharge due to their ability to retain moisture and promote infiltration through their root systems [105]. Agricultural land covers most of the basin (60.61%) and has a moderate to high recharge potential. The potential for over-extraction tempers the recharge benefit due to irrigation demands, which can lead to diminished groundwater levels if not carefully managed. Rangelands, primarily used for grazing, offer moderate recharge potential. Settlement areas with impermeable surfaces, such as asphalt and concrete, prevent water from penetrating, increasing surface runoff and decreasing aquifer replenishment [106].

4.4. Discussion

Aquifer Health Assessment

The aquifer health prediction map for the Barakar River Basin, developed using Random Forest, Decision Tree, and CNN models, categorises health into four classes: good, moderately good, semi-critical, and critical. In the Good category, CNN is the most optimistic, identifying 33.49% of the basin in this class, compared to 27.06% for Random Forest and 25.8% for Decision Tree (Figure 8). This suggests that CNN is more sensitive to features indicating favourable aquifer conditions. The Good health classification is mainly found in the upper catchment areas, including Mayurhand, Chauparan, Koderma, Domchanch, Padma, Markacho, and Deori. Decision Tree takes a more conservative approach for the Moderately Good category, designating 25.03% of the basin as moderately good. Random Forest and CNN, however, predict a smaller proportion, with 20.96% and 16.37%, respectively (Figure S1, see Supplementary File). These areas are spread across both the upper and lower catchments, notably, in Barhi, Ichak, Giridih, and Tundi. In the Semi-Critical category, Decision Tree again stands out, identifying the largest area (32.6%) as semi-critical, indicating a greater focus on areas at risk of aquifer stress. Both Random Forest and CNN show similar predictions for this category, with 24.33% and 24.04%, respectively. The Semi-Critical zones are predominantly located in the middle and lower catchments, especially in Jamua, Giridih, Tundi, Birni, and Nirsha. The Critical category shows the most distinct differences between the models. Random Forest is the most cautious, predicting 27.65% of the area as critical, highlighting concerns about potential aquifer depletion. CNN follows with 26.09%, while Decision Tree is more restrained, designating just 16.57% as critical. The Critical zones are primarily found in the lower catchment, particularly in the blocks of Narayanpur, Karma Tanr, Pirtanr, Bengabad, and Jamtara. CNN presents a more optimistic view of aquifer health, with a larger area classified as Good, whereas Random Forest is more conservative, marking a greater number of areas as Critical. The Decision Tree model offers a balanced approach, classifying regions as either Moderately Good or Semi-Critical.

The Pearson correlation coefficient analysis of the CNN-predicted aquifer health map, using 20 conditioning factors, reveals key insights into which factors most strongly impact aquifer health. Groundwater recharge exhibits the highest positive correlation (r = 0.52), making it the most influential factor in predicting the health of the CNN aquifer (Figure 9). This strong correlation suggests that areas with higher recharge rates are likely associated with better aquifer health, indicating the importance of sustainable recharge processes. While groundwater recharge plays a significant role in aquifer health, it is important to note that the quality of infiltrating water also influences groundwater conditions. While the model emphasises recharge rates as a key factor in aquifer health, the potential impact of poor-quality infiltrating water is acknowledged in the Water Quality Index (WQI) section. Return flow (r = 0.50) and groundwater fluctuation (r = 0.48) also exhibit strong positive correlations, underscoring their significant roles. These factors are crucial in the assessment, likely because they directly affect groundwater availability and stability. WQI (r = 0.38) and groundwater extraction for domestic purposes (r = 0.30) exhibit moderate positive correlations, indicating that higher water quality and controlled domestic extraction contribute to improved aquifer health. Other factors, such as aquifer media (r = 0.27), groundwater extraction for irrigation (r = 0.28), and land use/land cover (LULC) (r = 0.19), also show moderate to low positive correlations, implying that geological characteristics and land use patterns affect aquifer health, albeit to a lesser extent. Factors like precipitation (r = −0.34), lineament density (r = −0.34), and lateral flow (r = −0.33) show moderate negative correlations. These findings suggest that, in this context, increased precipitation, denser lineament features, and lateral water flow may not directly contribute to positive aquifer health, possibly due to issues like excessive runoff or low infiltration in certain areas. Other factors with weak negative correlations, such as surface runoff (r = −0.18) and base flow (r = −0.17), have a limited but notable influence on aquifer health. Less influential factors include pond density (r = 0.06) and groundwater extraction for industrial purposes (r = 0.08), which have minimal correlations, indicating that they do not significantly impact the predicted aquifer health in this region. These insights suggest that factors directly tied to groundwater dynamics and recharge potential are essential in assessing aquifer health. In contrast, some land use and surface-water indicators play a lesser role in this model’s context.

This study explicitly accounts for both groundwater quality degradation and quantity-related stressors in the classification of aquifer health. WQI, derived from key chemical indicators, contributed to identifying critical zones such as Koderma, Pirtanr, and Deori, where water quality was notably poor. The correlation analysis shows that groundwater recharge (r = 0.52) and return flow (r = 0.50) had stronger associations with aquifer health than WQI (r = 0.38), suggesting that quantity-related dynamics had a comparatively greater influence in the model’s predictions. The model does capture the dual impact of degraded water quality and overextraction, as seen in blocks with clean water but high extraction pressure being classified as “moderately good” or “semi-critical”. While water quality is an important factor, the model reflects a balanced consideration rather than placing exclusive or greater weight on quality impairments alone.

The study is grounded in a specific basin; however, the methodology itself is not region-specific. It integrates universally available datasets (e.g., remote sensing, climate reanalysis, publicly accessible geospatial layers) and scalable modelling tools (e.g., SWAT, machine learning, deep learning architectures), which have been successfully applied in various global contexts. The proposed hybrid modelling framework offers a robust, scalable approach to assess aquifer health across heterogeneous and data-scarce regions. While the methodology may appear complex, it provides distinct advantages by integrating surface hydrology, subsurface conditions, and anthropogenic factors into a unified assessment scheme. In large and diverse river basin settings, where field surveys are often constrained by time, cost, and logistical limitations, such an integrated model offers a practical alternative. This multi-dimensional approach improves the framework’s transferability, making it particularly suitable for semi-arid and hard-rock terrains, where conventional assessment methods often face limitations in spatial coverage and temporal frequency. The integration of machine learning techniques with hydrological models like SWAT has been successfully applied in similar regions, demonstrating their ability to predict and identify stress zones and water quality [13,107,108]. Deep learning models such as CNN have also shown high classification accuracy for groundwater health in semi-arid regions, where groundwater stress and contamination are significant concerns [109,110]. The application of geophysical techniques like Electrical Resistivity Tomography (ERT) improves subsurface characterisation and enhances model accuracy. Studies combining ERT with machine learning have demonstrated enhanced performance in delineating aquifer health zones and offer promising avenues for use in data-scarce and geologically complex environments [111,112]. These methods help identify critical aquifer zones, allowing for targeted interventions such as managed aquifer recharge (MAR) for groundwater protection. The framework’s adaptability to diverse geological settings, along with the growing demand for data-driven environmental governance, underscores its potential as a powerful tool for sustainable groundwater management.

While the complexity of hydrogeological systems in large river basins is acknowledged, this study deliberately focuses on shallow aquifers, the most exploited and vulnerable groundwater resources in the region. To ensure data reliability, both primary and secondary geospatial datasets were integrated, sourced from trusted platforms such as CGWB, GSI, NASA POWER, India-WRIS, and the FAO-UN Soil Database. Rigorous filtering and spatial validation were applied to maintain consistency and relevance to shallow aquifer conditions. Although some generalisation was necessary due to data limitations, the hybrid modelling approach, combining SWAT with machine learning and deep learning, offers a robust, scalable solution. This framework effectively balances scientific rigour and practical feasibility, serving as a valuable decision support tool for aquifer health assessment and groundwater management in similarly complex, data-scarce regions.

The results of this study have direct implications for water resource planning. The generated aquifer health maps can guide groundwater managers, policymakers, and environmental planners in prioritising zones for intervention. By identifying zones with different levels of groundwater stress and recharge, the maps enable targeted management actions, such as prioritising critical areas for MAR initiatives, regulating groundwater abstraction to prevent overexploitation, and guiding land use planning to control contamination sources in vulnerable zones. The methodology is adaptable to other basins experiencing comparable hydrogeological stress, thereby supporting evidence-based decision-making and promoting sustainable groundwater governance at a broader scale.

5. Validation

The validation results for the Decision Tree, Random Forest, and CNN models in assessing aquifer health in the Barakar River Basin offer insights into the performance of each model. Focusing on the AUC-ROC, a key metric for evaluating the model’s ability to discriminate between classes, the CNN achieves the highest AUC score of 0.95 (Figure 10). The diagonal dotted line in the ROC curve represents the performance of a random classifier (AUC = 0.5), serving as a baseline for comparison. The substantial deviation of all model curves above this line confirms their effectiveness. This high AUC indicates that the CNN model is highly effective in distinguishing between different health categories for aquifers, with a strong balance between true positive rate (TPR) and low false positive rate (FPR). Random Forest follows with an AUC of 0.859, also showing robust performance, though slightly less precise than CNN. The Decision Tree has the lowest AUC of 0.823, suggesting it may be less effective than the other models at differentiating between categories.

Precision, recall, and F1 score provide further insight into the models’ predictive capabilities (Table S5, see Supplementary File). CNN has the highest precision (0.957), indicating that it makes fewer false positive predictions, which is essential for minimising overestimation of aquifer health. However, its recall (0.729) is lower than that of both Random Forest and Decision Tree, suggesting that CNN may miss some true positives, which could mean some potentially critical areas might be under-identified. The F1 score, which balances precision and recall, is highest for CNN at 0.828, demonstrating a good overall performance with a slight bias towards precision. In comparison, the Random Forest model achieves a balanced precision (0.767) and recall (0.81), yielding an F1 score of 0.788. This suggests that Random Forest provides a more balanced approach between sensitivity (recall) and specificity (precision) compared to CNN. The Decision Tree has the lowest precision (0.752) and F1 score (0.78) among the models. However, its recall (0.81) matches that of the Random Forest, suggesting it identifies the majority of true positives, albeit at the cost of more false positives. CNN demonstrates superior discriminative ability with the highest AUC-ROC and precision, making it highly reliable for identifying non-critical zones. Random Forest, with a slightly lower AUC but balanced precision and recall, is effective at capturing critical areas with minimal compromise in accuracy. Though less effective overall, the Decision Tree still shows acceptable performance, particularly in recall. These variations highlight CNN’s potential as a precise model and Random Forest’s strength as a balanced classifier for aquifer health assessment in the Barakar River Basin.

The validation of the CNN-generated aquifer health assessment map in this study was achieved using Electrical Resistivity Tomography (ERT) surveys conducted at four strategically selected locations within the study area. These sites were chosen to represent both good aquifer health zones (P1: Parsauni in Itkhori block and P2: Madhopur in Dhanwar block) and critical aquifer health zones (P3: Mandro in Pirtanr block and P4: Phuljori in Jamtara block), as identified by the CNN-predicted map (Figure 11). The ERT results provided robust evidence in support of the predicted classifications. The resistivity profiles in the good aquifer health zones revealed shallow, water-saturated layers indicative of robust aquifer conditions. Profile 1 revealed a low-resistivity layer (~5–44 Ωm) extending to a depth of approximately 20 m, indicating a water-bearing zone beneath a soil layer with high moisture content. Below this, a sharp increase in resistivity (~1000–8000 Ωm) delineated the transition to a compact rock formation. Similarly, Profile 2 exhibited low to moderate resistivity (~2–80 Ωm) up to approximately 25 m in depth, consistent with a sustainable water-bearing formation, underlain by compact rock, as indicated by higher resistivity values (~300–1500 Ωm). These findings align with the CNN-generated map’s prediction of good aquifer conditions at these locations.

The profiles in the critical aquifer health zones revealed limited or deeper aquifer formations. Profile 3 predominantly exhibited high resistivity (~150–1500 Ωm) at shallow depths, indicating compact rock with minimal water-bearing capacity. However, a low-resistivity anomaly (~1–25 Ωm) detected at greater depths (70–100 m) suggests the presence of a confined water-saturated zone. Profile 4 similarly showed compact rock formations with high resistivity (~400–4500 Ωm) at depths of 10–15 m, accompanied by a localised low-resistivity anomaly (~15–30 Ωm) at approximately 35 m depth, indicating a limited aquifer. These observations confirmed the CNN map’s identification of these areas as critical aquifer zones where groundwater availability is constrained. The strong agreement between the CNN-predicted aquifer health zones and the ERT survey results demonstrates the reliability of the proposed methodology for assessing aquifer health.

Although only four ERT profiles were conducted, the sites were purposefully chosen to represent diverse geological conditions and contrasting aquifer health zones, enhancing the robustness of this supplementary validation. The close agreement between resistivity data and CNN predictions validates the model’s outputs, particularly for shallow aquifers, which are the primary focus of this study. Interpreting ERT results inherently requires consideration of the geological context, as resistivity values are influenced not only by water content but also by lithology and structural characteristics. In this study, interpretations of resistivity sections were made with reference to known geological conditions of the surveyed areas, including expected lithologies and existing hydrogeological insights. Zones of high resistivity were associated with compact or massive rock formations, while low-resistivity zones were attributed to weathered or fractured layers potentially saturated with water [113,114]. This integrative approach helped reduce ambiguity in interpretation and ensured that geophysical data were evaluated within an appropriate geological framework. We acknowledge the influence of lithology and structural complexity on resistivity interpretations; the ERT method remains well-suited for shallow subsurface characterisation. ERT served as an additional qualitative validation layer rather than the primary evaluation tool. The study’s main validation rests on statistical performance metrics. Deeper aquifer systems lie beyond the present scope but are recommended for inclusion in future investigations to expand the framework’s applicability to a broader hydrogeological context.

6. Conclusions

Combining tree-based categorisation, deep learning, and the SWAT model—a framework in a field of study where there is still a dearth of literature—offers an intriguing approach to aquifer health evaluation in the Barakar River Basin. The concept of aquifer health assessment itself is relatively new, with limited previous studies addressing this critical aspect of groundwater sustainability. This research fills a significant gap, as no prior work has combined machine learning, deep learning, and hydrological modelling for aquifer health assessment. The CNN model has emerged as the most effective classifier in this approach, achieving the highest AUC-ROC (0.95) and precision (0.957). This performance highlights CNN’s capability to accurately distinguish between aquifer health categories, particularly in non-critical zones, with minimal false positives. Non-critical zones refer to areas classified as Good or Moderately Good in aquifer health, indicating minimal groundwater stress and relatively stable recharge and extraction conditions. However, CNN’s recall suggests it may slightly under-detect some critical areas. The Random Forest model provides a balanced classification, with high recall (0.81) and precision (0.767), making it a dependable tool for identifying both healthy and stressed zones. Although the Decision Tree model has slightly lower performance, its high recall ensures that it captures a majority of actual positive cases, making it suitable for identifying critical areas, albeit at the cost of a few more false positives. Correlation analysis reveals the significance of factors such as groundwater recharge, return flow, and groundwater fluctuation, which exhibit strong positive correlations with the CNN model predictions and are crucial to the health of aquifers. The inverted resistivity results revealed shallow water-bearing zones in good health areas (P1 and P2), with resistivity values of ~5–44 Ωm and ~2–80 Ωm, while in critical regions (P3 and P4), limited aquifer formations were observed at greater depths, with resistivity values ranging from ~1 to 25 Ωm. These findings substantiate the CNN map predictions, as well as the ERT survey results, reinforcing the reliability of the proposed methodology. Although geological heterogeneity and deeper aquifer systems were not thoroughly investigated, the current study is concentrated on shallow aquifers. These elements suggest crucial avenues for further research that may improve this framework’s vertical and spatial applicability. In extensive and heterogeneous basins, where field surveys are hindered by time, cost, and logistical challenges, the integrated model presents a viable and efficient alternative.

This work not only advances the field of aquifer health assessment but also opens pathways for future research by demonstrating the value of integrating machine learning, deep learning, and hydrological models. This novel approach to aquifer health assessment holds significant global implications, providing a scalable framework for sustainable aquifer management in diverse regions facing similar water scarcity and quality challenges.

Take-home messages include the following:

• Integrated Approach: a novel hybrid framework combining SWAT, tree-based classification, and CNN effectively assesses aquifer health, addressing a gap in existing groundwater sustainability studies.
• Model Performance: the CNN model showed superior classification accuracy (AUC-ROC = 0.95), while Random Forest achieved high recall, ensuring balanced identification of both healthy and critical zones.
• Practical Utility: Electrical Resistivity Tomography (ERT) validation confirmed the CNN-based aquifer health maps, demonstrating the approach’s reliability and potential for replicable groundwater management in semi-arid, hard-rock terrains.