Integration of geospatial techniques and artificial neural networks for groundwater potential zonation in India

Abstract

This study delineates Groundwater Potential Zones (GWPZs) in the Mahi River Basin (MRB), where increasing groundwater stress is driven by urbanization, intensive agriculture, and limited recharge. Ten predictive factors: geology, geomorphology, slope, lineament density, drainage density, land use/land cover (LULC), Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), Normalized Difference Water Index (NDWI), and Toposoil Grain Size Index (TGSI), were derived from multisource satellite imagery (Landsat 8 OLI/TIRS, Sentinel-2, SRTM DEM) and field observations. An Artificial Neural Network (ANN) model assigned nonlinear contribution weights to these inputs and generated a GWPZ map with five classes: Very Good (5.24%), Good (16.35%), Moderate (29.89%), Poor (30.88%), and Very Poor (17.64%). Model performance was validated quantitatively using the Area Under the ROC Curve (AUC = 0.850). Additional validation using 350 well locations showed strong spatial correspondence between predicted zones and groundwater depth. The Kruskal–Wallis H test (H = 42.87, p < 0.001) confirmed significant hydrogeological differences among ANN-predicted classes, while Moran’s I (0.037, p = 0.46) demonstrated spatial independence of prediction residuals. More than 62% of the basin falls within Moderate to Very Poor zones, emphasizing the need for strategic groundwater recharge planning. The integration of ANN with remote sensing and GIS offers a robust, data-driven model for sustainable groundwater resource management in semi-arid and data-poor regions.

Introduction

Groundwater is one of the most critical natural resources on Earth, playing a pivotal role in sustaining life and livelihoods by meeting freshwater demands across domestic, agricultural, and industrial sectors¹. Its significance is especially pronounced in regions with limited or unreliable surface water availability, such as arid and semi-arid zones, where groundwater often serves as the primary or sole source of water². Over the past few decades, escalating demands driven by rapid population growth, urbanization, agricultural intensification, and land use changes have placed immense pressure on groundwater systems. This has led to unsustainable extraction rates, surpassing the natural recharge capacity of aquifers. As a result, a continuous decline in groundwater levels is being observed, threatening not only long-term water security but also ecosystem balance, agricultural productivity, and overall environmental sustainability^3–6. Addressing this growing imbalance has become a critical priority in achieving sustainable water management, particularly in climate-stressed and resource-scarce regions.

India, in particular, faces significant challenges related to groundwater depletion⁷. Many regions are already experiencing water stress, and with projections indicating increased demand and climate variability, the pressure on groundwater is expected to intensify. In this context, systematic and scientific delineation of Groundwater Potential Zones (GWPZs) is vital for sustainable water management. Effective GWPZ mapping supports targeted recharge interventions, informed policy decisions, and optimized resource use^8–10.

Modern advancements in geospatial technologies - such as remote sensing (RS) and geographical information systems (GIS) - have made it possible to assess GW potential with greater accuracy and efficiency¹¹. These tools enable the extraction and integration of spatially distributed variables like geology, geomorphology, slope, land use/land cover (LULC), drainage and lineament densities, and vegetation indices (NDVI, SAVI, NDWI, TGSI). When these factors are analyzed together, they comprehensively represent surface and subsurface conditions influencing groundwater dynamics^12,13.

To move beyond traditional, linear-based weighting techniques, this study employs an ANN-based model, which simulates the learning behavior of the human brain to detect nonlinear patterns and complex interrelationships among contributing parameters^14,15. ANN models can refine predictions through iterative training, offering a data-driven, adaptive, and highly accurate approach to GWPZs mapping^16–19. This methodology is particularly valuable in areas with limited subsurface data, as it leverages observable surface features and historical records to infer groundwater conditions. By continuously adjusting internal weights based on error minimization, ANN systems improve classification performance with each training cycle. Moreover, their capacity to handle non-linear relationships among multiple thematic layers makes them especially suitable for complex hydrogeological terrains^20–22.

Several studies have highlighted the effectiveness of ANN in GWPZ mapping. For instance, ANN models have been successfully applied to delineate GWPZs with higher prediction accuracy compared to traditional methods. The applicability of ANN for groundwater vulnerability assessment has also been demonstrated in diverse hydrogeological settings^23–25. Additionally, the integration of ANN with remote sensing and GIS has proven effective in managing complex spatial datasets and capturing nonlinear relationships between variables^26,27. These studies collectively emphasize the robustness of ANN as a machine-learning tool in hydrogeological applications, especially in data-scarce regions. Considering the challenges of groundwater assessment in semi-arid areas, ANN offers a practical and scalable solution for identifying potential recharge zones. The present study builds upon this foundation by adapting ANN techniques to local conditions, thereby contributing to informed water resource management.

This study addresses the escalating groundwater stress in the MRB, a region experiencing rapid environmental and socioeconomic changes. The novelty of the research lies in the integration of multi-criteria spatial datasets with ANN modeling to delineate GWPZs with greater precision than conventional approaches. Unlike earlier studies that rely primarily on AHP or index-based methods, this research leverages the pattern recognition and non-linear modeling capabilities of ANN to enhance prediction accuracy. The generated GWPZ map is validated using both well inventory data and groundwater depth observations, ensuring scientific rigor and practical relevance. The primary contribution of this work is the development of a high-resolution, reliable GWPZ model using Remote Sensing, GIS, and ANN techniques. This approach not only supports groundwater exploration and recharge planning within MRB but also offers a scalable and transferable framework for sustainable groundwater management in other data-scarce or semi-arid river basins globally, contributing to broader goals of water security and environmental sustainability.

Materials and methodology

Study area

The MRB is located in western India and spans parts of Madhya Pradesh, Rajasthan, and Gujarat, ultimately draining into the Arabian Sea through the Gulf of Khambhat (Fig. 1). The Mahi River originates in Madhya Pradesh and is fed by three principal tributaries: the Som River, rising in the Eastern Aravalli Hills of Rajasthan; the Anas River, originating in the South Vindhyan Ranges of Madhya Pradesh; and the Panam River, which joins the Som and Anas in the Durgapur region of Rajasthan^28,29. These tributaries contribute significantly to the basin’s drainage dynamics and hydrological behavior³⁰. The MRB exhibits diverse physiographic features, including the rugged Eastern Aravalli Hills, undulating Vindhyan terrain, and intervening plains, resulting in complex geomorphological and hydrological characteristics. The region’s semi-arid climate, marked by distinct seasonal variations in precipitation and temperature, poses significant challenges to groundwater recharge and availability³¹. Moreover, the basin is characterized by high lineament density, variable drainage patterns, and heterogeneous geological formations, all of which influence subsurface water movement and accumulation. Land use in the MRB is predominantly agricultural, with increasing urbanization exerting additional pressure on groundwater resources. Population density varies across the basin, with urban centers witnessing rapid growth, while rural communities rely heavily on groundwater for irrigation and domestic use. The high lineament density in certain zones, indicative of enhanced subsurface permeability and groundwater movement combined with increasing anthropogenic pressure, highlights the urgent need for a scientific assessment and sustainable management of groundwater resources in the region³².

Fig. 1.

Study area map of the MRB showing geographical extent, elevation (m), and streams extracted from the DEM. The map was generated in ArcMap 10.8 (https://www.esri.com/en-us/arcgis/products/arcgis-desktop/overview) using SRTM DEM downloaded from USGS Earth Explorer (https://earthexplorer.usgs.gov/) and Survey of India shapefiles (https://onlinemaps.surveyofindia.gov.in/Digital_Product_Show.aspx). The final map was overlain on an OpenTopography basemap within ArcGIS to enhance terrain visualization.

Data used

This study employed multisource satellite and collateral data to delineate GWPZs in the MRB. Satellite imagery was acquired in 2024, sourced from Landsat 8 OLI/TIRS (30 m spatial resolution) and Sentinel-2 MSI (10 m spatial resolution). A Digital Elevation Model (DEM) with a 30 m spatial resolution was obtained from Earth Explorer (https://earthexplorer.usgs.gov/), while Sentinel-2 data were accessed through the Copernicus Open Access Hub (https://scihub.copernicus.eu/).

The DEM enhanced and improved accuracy; preprocessing steps such as resampling and noise reduction were applied, followed by integration with 1:50,000 scale topographic sheets obtained from the Survey of India (SOI). These sheets provided detailed elevation and terrain features, supporting refinement of topographic and hydrological parameters. Collateral datasets, including meteorological and climate information, were retrieved from the Meteorological and Oceanographic Satellite Data Archival Centre (MOSDAC). Fieldwork was conducted using a high-precision Trimble Geo 7X GPS receiver to collect accurate latitude and longitude coordinates. These ground control points (GCPs) were essential for georeferencing, spatial alignment, and validation of geospatial layers.

Image calibration and atmospheric correction were performed to ensure data integrity and enhance the reliability of subsequent analysis. Geometric corrections were conducted using ArcGIS 10.8, where GCPs and polynomial transformations (second-order) were applied to minimize spatial distortions. The Root Mean Square Error (RMSE) was calculated to confirm geometric accuracy, with acceptable thresholds maintained throughout. Radiometric corrections included noise reduction techniques such as dark object subtraction and histogram equalization to compensate for sensor-specific distortions and haze effects. Atmospheric corrections, including haze and cloud removal, were performed using ERDAS IMAGINE. Additionally, resampling methods (cubic convolution) were employed to ensure uniform spatial resolution across datasets. The processed images were integrated within a GIS environment to support comprehensive spatial analysis. Table 1 summarizes the data sources, spatial resolutions, acquisition platforms, and preprocessing techniques, including noise reduction, resampling, and atmospheric corrections used for mapping GWPZs in the MRB, thereby ensuring robust and reliable results.

Table 1.

Data type, Sources, and their Spatial resolutions.

Data type	Source	Spatial resolution	Collection year	Purpose/description
Landsat 8 OLI/TIRS	Earth Explorer (https://earthexplorer.usgs.gov/)	30 m	2024	Used for generating thematic maps.
Sentinel-2	Copernicus Open Access Hub (https://scihub.copernicus.eu/)	10 m	2024	Provides high-resolution imagery for detailed analysis of land cover/land use.
Digital elevation model (DEM)	Earth Explorer (https://earthexplorer.usgs.gov/)	30 m	2024	Used to understand topography and elevation changes; enhanced with resampling and noise reduction.
Topographic sheets	Survey of India (SOI)	1:50,000 scale	2024	Incorporated to validate and refine DEM data; provides detailed terrain and elevation information.
Meteorological and climate data	Meteorological and Oceanographic Satellite Data Archival Centre (MOSDAC)	Varies	2024	Provides contextual climate and weather information essential for analysis and interpretation.
GPS coordinates	Trimble Geo 7X	High precision	–	Collected for accurate georeferencing and alignment of datasets; ensures spatial accuracy.
Image rectification	ArcGIS 10.8 software	–	–	Corrects geometric distortions and aligns imagery to a standard coordinate system using ground control points (GCPs).
Data processing software	ERDAS IMAGINE 2014 and ArcGIS 10.8	–	–	Used for comprehensive data processing, analysis, and interpretation of geospatial data.

Methodology

This study employed a data-driven approach using ANN integrated with GIS to delineate GWPZs in the MRB, India. The methodology consists of three core stages: thematic layer preparation, ANN model development and training, and GWPZ mapping and validation. The overall workflow is illustrated in Fig. 2.

Fig. 2.

Methodology flowchart. the flowchart was prepared using MS office powerpoint.

Satellite data acquisition and preprocessing

Landsat 8 OLI/TIRS satellite imagery was procured and atmospherically corrected using standard preprocessing techniques. Various thematic indices were generated using band combinations and raster operations via the Raster Calculator tool in the Spatial Analyst toolbox of ArcGIS. Table 2 presents the band ratio equations that derive remote sensing-based thematic layers.

Table 2.

Band rationing equations for (Spectral indices ) thematic Layers.

Thematic map	Band combination	Landsat 8 OLI bands	Equation	References
NDVI	NDVI = (NIR - R) / (NIR + R)	NDVI = (Band 5 – Band 4) / (Band 5 + Band 4).	(1)	33,34
SAVI	SAVI= ((NIR - R) / (NIR + R + L)) * (1 + L)	SAVI = ((Band 5 – Band 4) / (Band 5 + Band 4 + 0.5)) * (1.5).	(2)	35
NDWI	NDWI = (NIR – SWIR) / (NIR + SWIR)	NDWI = (Band 5 – Band 6) / (Band 5 + Band 6)	(3)	36,37
TGSI	TGSI= (Red-Blue)/(Red + Blue + green)	TGSI = (Band 4 – Band 2) / (Band 4 + Band 2 + Band 3)	(4)	38,39

Selection, preparation, and correlation analysis of thematic layers

Thematic layers (TLs) representing environmental and hydrogeological parameters were selected based on their significance in influencing groundwater dynamics. These include geology, geomorphology, slope, lineament density, drainage density, land use/land cover (LULC), and remote sensing indices such as the NDVI, SAVI, NDWI, and TGSI.

Each layer was prepared using standard remote sensing and GIS-based procedures. Geology and Geomorphology were digitized from Geological Survey of India (GSI) maps and classified using Landsat 8 imagery. Drainage and Drainage Density were derived from the Digital Elevation Model (DEM) using the ArcHydro tool. Slope was generated from the DEM using the “Slope” function in ArcGIS 10.8. LULC was classified from Sentinel-2 data using the supervised classification method. Lineament Density was extracted from satellite and DEM data using the Line Density tool. Vegetation and Surface Indices, NDVI, SAVI, NDWI, and TGSI were computed from Landsat 8 band ratios.

All layers were rasterized, resampled to a uniform 30 m spatial resolution, and normalized using min - max scaling to standardize the range of input features before ANN processing.

To ensure statistical independence among the selected thematic layers, a Pearson correlation matrix and Variance Inflation Factor (VIF) analysis were performed⁴⁰. The results (Table 3) show that all correlation coefficients were below 0.80 and VIF values were under 5, confirming the absence of significant multicollinearity among predictors⁴¹. Although NDVI and SAVI exhibited a moderate correlation (r = 0.72), both were retained because they represent distinct surface characteristics relevant to groundwater potential NDVI captures vegetation vigor, while SAVI adjusts for soil brightness effects. These findings validate that each thematic layer independently contributes to the ANN model performance.

Table 3.

Correlation matrix and variance inflation factors (VIFs) for predictor variables used in the ANN Model.

Variable	Geology	GM	Slope	DD	LD	NDVI	SAVI	NDWI	TGSI	LULC	VIF
Geology	1	0.3	0.21	0.3	0.2	0.09	0.08	0.12	0.11	0.19	2.1
GM		1	0.36	0.3	0.3	0.14	0.15	0.17	0.18	0.22	2.4
Slope			1	0.4	0.3	−0.18	-0.19	−0.15	−0.17	−0.25	2.8
DD				1	0.4	−0.16	-0.17	-0.12	−0.14	−0.21	3.1
LD					1	0.09	0.08	0.11	0.12	0.15	1.9
NDVI						1	0.72	0.35	0.38	0.48	3.8
SAVI							1	0.33	0.36	0.47	3.9
NDWI								1	0.29	0.35	2.6
TGSI									1	0.22	2.3
LULC										1	3

All correlation coefficients (|r| < 0.80) and VIF values (< 5) indicate no significant multicollinearity among predictor variables. Although NDVI and SAVI are moderately correlated, both were retained since NDVI captures overall vegetation vigor and SAVI compensates for soil brightness effects both crucial for hydrological and recharge modeling.

Ground truth data collection

High-yield well locations, including borewell data obtained from the Central Water Board (CWB) database, were compiled and spatially aligned with the raster stack of thematic layers. Each location was categorized into groundwater yield classes (High, Moderate, Low) based on measured water level data. This dataset was used as the labeled input for supervised ANN model training.

ANN model architecture and training

An ANN-based classification model was developed using the Weka machine learning platform, with spatial datasets pre-processed in ArcGIS 10.8^42–44. The raster stack of thematic layers was converted into tabular form and exported as attribute data for ANN processing in Weka. The input layer size corresponded to the number of thematic layers used as predictive features. Two hidden layers were configured with 32 and 16 neurons, respectively, employing ReLU activation functions to introduce non-linearity^45,46. The output layer included five neurons with a Softmax activation function, directly classifying the data into Very Good, Good, Moderate, Poor, and Very Poor GWPZs, following standard multi-class classification practice as used in similar ANN studies^48,49. The contribution (importance) of each thematic layer was calculated using the Garson algorithm, which partitions the connection weights between input, hidden, and output layers to quantify the relative influence of each predictor within the ANN model.

To improve generalizability and reduce overfitting risks, the dataset was split into training (70%), validation (15%), and testing (15%) subsets. Early stopping was applied to terminate training when validation accuracy failed to improve over 15 consecutive epochs. Dropout regularization (rate = 0.2) was introduced in the hidden layers to randomly deactivate neurons during training, thereby enhancing robustness. Model training followed a supervised learning approach, using categorical cross-entropy as the loss function and the Adam optimizer for weight updates.

Training parameters included a batch size of 32, 100 epochs, and a validation split of 30% of the input dataset for evaluating generalization performance. Similar ANN-based groundwater potential studies, used architectures with one or two hidden layers and neuron counts ranging from 10 to 30, achieving accuracies between 94% and 85%^49,50. In comparison, the architecture adopted in this study achieved an accuracy of 85%, demonstrating competitive performance and robustness in mapping GWPZs under semi-arid conditions.

In this study, multiple thematic layers including slope, geology, geomorphology, lineament density, drainage density, NDVI, SAVI, NDWI, TGSI, and land use/land cover (LULC) were utilized as predictive factors. Each layer represents unique spatial and environmental characteristics influencing groundwater occurrence and recharge potential. While many of these factors capture distinct hydrogeological or surface conditions, some, such as NDVI and SAVI, are highly correlated as both are related to vegetation cover. NDVI measures vegetation health and greenness, whereas SAVI adjusts for soil brightness effects, especially in areas with sparse vegetation. Despite this correlation, both indices offer complementary insights, with NDVI focusing on vegetation vigor and SAVI enhancing signal detection where soil background impacts vegetation reflectance. Therefore, both were retained as independent input features to enrich the model’s ability to discern subtle variations in surface conditions that affect infiltration and recharge.

Other factors such as slope, which controls runoff and infiltration patterns; geology and geomorphology, which influence aquifer structure and storage capacity; lineament density, indicating secondary porosity pathways; drainage density, which affects surface runoff; NDWI, capturing moisture presence; TGSI, reflecting topsoil greenness; and LULC, determining land surface characteristics, were also incorporated as independent variables. Collectively, these factors provide a holistic representation of the surface and subsurface environment, and their inclusion ensures that the ANN model effectively captures diverse processes governing groundwater dynamics.

Model evaluation

Model performance was evaluated using accuracy, which measures the overall correctness of the predictions. Additionally, the Receiver Operating Characteristic (ROC) curve was employed to visualize the model’s classification performance, while the Area Under the Curve (AUC) was used to quantify its discriminative ability, with higher AUC values indicating better model performance^51,52.

GWPZ mapping and visualization

After successful training, the model was used to predict GWPZs for the entire study area. For each pixel in the study area, the corresponding values from the normalized thematic layers were input into the ANN to classify the zone (Very good, Good, Moderate, Poor, and Very poor GWPZs). The resulting classification map was post-processed in ArcGIS to visualize spatial distributions.

Validation with field data

The final GWPZ map was validated using overlay analysis of well yield points on the predicted GWPZs to assess spatial agreement⁵³. A comparison with observed water table depth data was also performed to ensure consistency with actual hydrogeological conditions. Model performance was quantified using the AUC-ROC curve and a confusion matrix, which provided class-wise precision, recall, and F1-scores. To verify statistical differentiation among groundwater potential classes, a Kruskal–Wallis H test with Dunn’s post-hoc correction was performed, confirming significant differences (p < 0.001) in groundwater depths across the five GWPZs. Additionally, a Moran’s I test and residual semivariogram analysis were applied to evaluate spatial independence of prediction errors. The results (I = 0.037, p = 0.46) indicated random residual distribution, confirming that the ANN model is free from spatial bias.

Advantages of ANN over traditional AHP

The ANN technique was preferred over the AHP due to its data-driven, objective nature. While AHP depends on expert judgment for assigning weights introducing subjectivity, ANN autonomously learns from data, capturing complex nonlinear relationships among variables. ANN offers better consistency, scalability, and adaptability, especially across large and geologically diverse areas. Its superior performance in recent studies further supports its use for more accurate and robust GWPZ mapping in this study.

ANN-based results and thematic layer influence

Feature importance and thematic contribution analysis

The trained ANN model revealed varying degrees of influence among the input thematic layers used for GWPZ classification. The model extracted relative importance scores (weights) assigned to each input feature during training using built-in feature analysis techniques. Feature-importance values were computed using the Garson algorithm, which decomposes connection weights across input, hidden, and output layers to quantify the relative contribution of each thematic variable. These scores indicate how significantly each thematic layer contributes to the classification of GWPZs. Table 4 presents the summary statistics of each thematic layer, including the number of categories and their respective ANN-derived contribution weights. The highest weight was observed for Lineament Density (14%), followed by Geomorphology (12%), Drainage Density (12%), Slope (11%), and LULC (10%). Remote sensing-based indices such as SAVI, NDVI, and NDWI demonstrated moderate importance, whereas TGSI showed a comparatively lower impact.

Table 4.

ANN-Based GWPZ categories with Spatial Extent, depth Ranges, and well data Overlay.

SL no.	GWPZs	Area (km2)	Area (%)	Depth to water level (m bgl)	No. of wells
1	Very Good	1,360.54	4.23	15.24 to 19.81	15
2	Good	5,541.25	17.21	20.12 to 24.38	60
3	Moderate	10,123.27	31.45	24.69 to 33.53	110
4	Poor	10,575.88	32.85	33.83 to 41.15	115
5	Very Poor	4,590.62	14.26	> 41.15	50
	Total	32,191.56	100		350

Thematic layers and their influence on GWPZs (ANN-based)

The ANN-based approach effectively quantifies the contribution of each thematic layer in delineating GWPZs. Rather than relying on subjective weights, ANN utilizes pattern recognition and training algorithms to extract underlying nonlinear relationships between the input thematic variables and groundwater occurrence⁵⁴. This method enhances precision and data-driven objectivity in understanding the roles of different parameters. Each thematic layer - ranging from geophysical properties to vegetation indices- impacts GWPZs directly, indirectly, or in a supporting manner. Based on the trained ANN model, the contribution weight of each layer was computed, considering the spatial and statistical relationship of its classes with groundwater occurrence. Table 4 below presents each thematic layer’s key statistics, source, number of classified categories, ANN-derived contribution weight, influence type, and impact on groundwater availability. Additionally, dominant influence classes are noted to highlight the key categories contributing most significantly to potential zones.

Table 4.

Detailed thematic layer statistics, classification ranges, dominant influence classes, and ANN-derived contribution weights for GWPZ mapping.

S. no.	Thematic layer	Contribution weight (%)	Number of classes	Class details (retained from original)	Dominant influence class (as per ANN)	Influence type	Description of impact on GWPZs
1	Geology	9	17	Acidic Bodies, Basic Sills/Dykes, Benpal Gp, Champaner, Coastal Sediments, Debari, Deccan Traps, Godhra Granite, Hindoli Gp, etc.	Deccan Traps	Direct	Lithological structure controls aquifer potential and permeability
2	GM	12	13	Alluvial Plain, Coastal Plain, Dam/Reservoir, Flood Plain, Dissected Hills/Valleys, Plateau, Pediment, Quarry, Waterbody types	Alluvial Plain	Direct	Controls infiltration, storage, and recharge potential
3	Lineament Density	14	5	0–0.15, 0.15–0.36, 0.36–0.57, 0.57–0.94, 0.94–1.78	0.94–1.78 (Very Fine)	Direct	Enhances secondary porosity and permeability
4	Drainage Density	12	5	0–1.02, 1.02–1.41, 1.41–1.75, 1.75–2.15, 2.15–3.26	0–1.02 (Very Coarse)	Indirect	Low drainage favors infiltration and reduces surface runoff
5	NDVI	7	5	−0.127–0.085, 0.085–0.125, 0.125–0.171, 0.171–0.255, 0.255–0.498	0.255–0.498	Supporting	Higher NDVI indicates healthy vegetation and potential recharge areas.
6	SAVI	8	5	−0.191–0.128, 0.128–0.187, 0.187–0.257, 0.257–0.383, 0.383–0.748	0.383–0.748	Supporting	Indicates vegetation condition influencing recharge indirectly
7	NDWI	6	6	−0.472–0.254, −0.254–0.197, −0.197–0.157, −0.157–0.074, −0.074–0.213, 0.101–0.420	0.101–0.420	Supporting	Moisture presence reflects surface wetness, aiding recharge.
8	TGSI	5	5	−0.132–0.047, 0.047–0.073, 0.073–0.088, 0.088–0.101, 0.101–0.420	0.101–0.420	Supporting	Surface greenness suggests favorable surface infiltration conditions.
9	Slope	11	5	0–2%, 2–5%, 5–15%, 15–35%, > 35%	0–2% (Very gently sloping)	Indirect	Gentle slopes promote percolation and reduce erosion
10	Land use/land cover (LULC)	10	6	Waterbody, dense vegetation, vegetation, fallow land, barren land, scrub	Vegetation	Supporting	Influences infiltration based on surface cover type

Geology

The geology of the MRB is notably diverse, comprising the Deccan basalts, the Aravalli Supergroup, and glacial deposits along the coast (Fig. 3(a)). The Aravalli Supergroup encompasses multiple formations, including the Hindoli Group, Mangalwar Complex, Pur-Banera Belt, Rajpura-Dariba Belt, Debari Group, Rakhabdev Ultrabasic Suite, Champaner Series, and Kankroli Group, each exhibiting unique lithological and structural characteristics that influence groundwater behavior⁵⁵. The Hindoli Group is predominantly composed of graywacke and phyllite, which offer moderate potential for groundwater infiltration due to their fine-grained texture and partial weathering. The Mangalwar Complex is primarily composed of gneissic rocks intersected by lineaments, which, despite their crystalline nature, enhance secondary porosity and allow limited water percolation. Similarly, the Pur-Banera Belt, characterized by Banded Iron Formations (BIF), quartzite, and conglomerates, provides potential recharge zones along fractured contacts. The Debari Group, with its petromict conglomerate and pebbly arkose, presents localized zones of moderate permeability. The Rakhabdev and Champaner Series, with varied lithologies including shale, limestone, and feldspathic quartzite, also contribute to groundwater potential, though limited by primary porosity. In the western part of the basin, the Deccan Traps dominate the geological setting. These Cretaceous-Paleocene volcanic rocks are typically impermeable, but their weathered and fractured zones can locally support groundwater recharge. The Godhra Granite, a Neoproterozoic formation underlying the Deccan Traps, further contributes to the basin’s complex hydrogeological character. The Semri Group, positioned in the northeastern part of the basin, is bordered by the Deccan Traps and Benpal Group. It includes formations of the Banded Gneiss Complex and Bundelkhand Granite, which form the basement of the Mesoproterozoic sedimentary sequences. The Benpal Group, of Archaean age, comprises amygdaloidal metabasalts and basic sills and dykes, which are typically complex and dense but may develop secondary porosity in weathered conditions.

Fig. 3.

Thematic layers used in the study. (a) Geology (digitized in ArcGIS using Bhukosh GSI maps), (b) Geomorphology (digitized in ArcGIS using Bhukosh GSI maps), (c) Lineament Density (derived from NRSC lineament data and processed in ArcGIS), and (d) Drainage Density (generated from SRTM DEM using ArcGIS). All datasets were processed in ArcMap 10.8 (https://www.esri.com/en-us/arcgis/products/arcgis-desktop/overview), with source maps obtained from Bhukosh GSI (https://bhukosh.gsi.gov.in/) and SRTM DEM from USGS Earth Explorer (https://earthexplorer.usgs.gov/). The final thematic layers were overlain on an OpenTopography basemap within ArcGIS to enhance spatial visualization and terrain contextualization.

In the ANN-based analysis, the geology thematic layer was assigned a contribution weight of 10%, based on the model’s learning from input patterns across the region. The ANN model identified regions underlain by weathered metamorphic rocks, sandy lithologies, and coarse-grained sediments that exhibit greater groundwater potential due to higher infiltration and storage capacity. In contrast, due to their limited porosity and permeability, massive crystalline rocks such as granites, basalts, and ultramafic units exhibited lower contributions. As such, geology provides the foundational framework for the ANN model to delineate areas favorable for groundwater storage and recharge, shaping the spatial pattern of GWPZs across the study area.

Geomorphology (GM)

Geomorphology plays a significant role in determining groundwater potential, as it directly influences surface runoff, infiltration, and subsurface water storage⁵⁶. In this study, geomorphic units were delineated from remote sensing data and categorized into standard landform classes, including alluvial plains, coastal plains, flood plains, dissected hills/valleys, plateaus, pediments, quarries, and waterbodies (Fig. 3(b)).

Pediment - pediplain complexes, alluvial plains, and flood plains emerged as high-influence classes in the ANN model due to their relatively flat terrain, weathered material cover, and greater infiltration capacity conditions favorable for groundwater recharge. Similarly, areas adjacent to rivers, canals, and reservoirs exhibited high groundwater potential due to sustained moisture availability and hydrological connectivity. In contrast, hard rock terrains, including dissected hills and plateaus, exhibited lower potential because of steep gradients, limited soil cover, and high surface runoff, which collectively limit recharge and storage. The ANN model assigned a 12% contribution weight to geomorphology, highlighting its substantial role in groundwater potential distribution. To address differences in hydrogeological behavior between alluvial and hard rock terrains, model performance was also assessed separately for these two major geomorphic settings. This analysis showed higher predictive accuracy in the alluvial plains, where aquifer continuity and recharge mechanisms are more uniform, compared to hard rock zones, which display higher variability due to fractured aquifer systems. The results underscore the importance of terrain-specific validation to ensure robust and context-sensitive groundwater potential mapping.

Drainage density (DD)

Drainage density (DD) is a critical morphometric parameter that reflects the extent and distribution of surface drainage features in a given area. It is calculated as the total length of streams per unit area and is inversely related to surface permeability and groundwater recharge^57,58. In this study, DD was classified into five ranges: 0 to 0.02 km/km², 0.02 to 0.06 km/km², 0.06 to 0.10 km/km², 0.10 to 0.14 km/km², and 0.14 to 0.26 km/km² (Fig. 3(c)). In the ANN model, areas with low drainage density - specifically the 0 to 0.02 km/km² class - were recognized as the dominant influence zones contributing positively to groundwater potential. These regions typically allow greater infiltration and lower surface runoff, facilitating water percolation into subsurface layers. In contrast, higher drainage density zones (0.14 to 0.26 km/km²) were learned by the ANN model to be less favorable for groundwater accumulation, as they are often associated with impermeable surfaces, steep gradients, and enhanced surface runoff - factors that reduce groundwater recharge efficiency. ANN-based approach assigned a 9% contribution weight to the drainage density layer, reflecting its indirect but significant influence on groundwater potential. The model effectively captured spatial patterns where stream frequency impacted recharge dynamics without relying on predefined human-assigned rankings.

Lineament density (LD)

Lineaments, encompassing geological structures such as faults, joints, and fractures, serve as critical pathways for groundwater movement and storage. These linear features are prominently detectable through satellite imagery and manifest as straight or curvilinear alignments across the terrain. Their presence often correlates with enhanced secondary porosity and permeability, making them vital indicators in groundwater potential assessments⁵⁹. The current study used Lineament Density (LD) as a key thematic layer in delineating GWPZs. The LD map was derived from remote sensing and DEM sources and categorized into five distinct classes: 0–0.15 km/km², 0.15–0.36 km/km², 0.36–0.57 km/km², 0.57–0.94 km/km², and 0.94–1.78 km/km² (Fig. 3(d)). These ranges were selected to represent increasing levels of structural complexity and groundwater conductivity. ANN-based analysis revealed that areas with the highest lineament density (0.94–1.78 km/km²) exhibited dominant influence in the model output, strongly correlating with high groundwater potential. This reinforces the hydrological significance of densely fractured and faulted zones as preferential pathways for recharge and subsurface water flow. Consequently, LD was assigned a 14% contribution weight - the highest among all factors - highlighting its strong predictive importance in ANN modeling for GWPZ identification.

NDVI

Vegetation is crucial in regulating groundwater dynamics, primarily through its influence on infiltration, evapotranspiration, and surface runoff⁶⁰. The NDVI is a widely used remote sensing index that captures the health and density of vegetation based on its spectral response. Vegetated surfaces typically absorb visible light (especially red wavelengths) and strongly reflect NIR radiation, a contrast that NDVI exploits to quantify vegetation cover and vigor⁶¹. In this study, NDVI was employed to assess vegetation’s spatial distribution and density across the region, offering insights into areas that favor groundwater recharge. The NDVI values derived for the study area range between − 0.127 and 0.498, where negative or near-zero values indicate barren land, urban areas, or water bodies, and higher positive values reflect denser and healthier vegetation (Fig. 4(a)). Vegetated surfaces, particularly those with dense canopy cover, improve water retention and enhance infiltration by reducing the impact of raindrops on the soil surface and lowering the velocity of overland flow.

Fig. 4.

Thematic layers used in the study (continued). (a) NDVI, (b) SAVI, (c) NDWI, and (d) TGSI. All spectral indices were generated and processed in ArcMap 10.8 ( https://www.esri.com/en-us/arcgis/products/arcgis-desktop/overview) using Landsat-8 OLI/TIRS satellite imagery downloaded from USGS Earth Explorer (https://earthexplorer.usgs.gov/). The final index layers were overlaid on the OpenTopography basemap within ArcGIS to enhance surface terrain interpretation and spatial visualization.

The thematic NDVI layer was integrated into the ANN-based analysis, contributing a weight of 9% to the overall model. This contribution was determined by the ANN’s capacity to recognize patterns and correlations between vegetation density and subsurface recharge potential. Areas with NDVI values closer to the upper end of the range were found to align with zones of higher groundwater potential due to their favorable surface conditions for percolation. Conversely, regions with low or negative NDVI values, representing either developed land or exposed surfaces, demonstrated minimal groundwater recharge capability. By including NDVI in the ANN framework, the model effectively captures vegetation’s indirect yet significant role in groundwater sustainability. This thematic layer enhances the predictive accuracy of the GWPZ mapping by incorporating ecological and surface conditions that influence the hydrological cycle at the ground level.

SAVI

The SAVI is instrumental in distinguishing water bodies, vegetative cover, and soil characteristics while mitigating the influence of soil brightness on vegetation measurements (Fig. 4(b)). By integrating a soil adjustment factor into its calculation, SAVI enhances the precision of vegetation assessment, particularly in areas with sparse or heterogeneous ground cover. This makes it a valuable indicator in hydrological studies, especially when evaluating surface permeability and the potential for groundwater recharge. In the current analysis, SAVI plays a significant role in assessing the extent and type of surface vegetation, which in turn impacts the infiltration capacity of the land⁶². Surfaces with dense vegetation facilitate better moisture retention and reduce surface runoff, promoting groundwater recharge.

On the other hand, areas with sparse vegetation or bare soil are more prone to erosion and runoff, limiting the amount of water that percolates into the subsurface. The SAVI values in the study area were computed and classified into five distinct ranges to reflect their hydrological significance: −0.191 to 0.128 corresponds to fine soil textures and scrublands, typically associated with poor infiltration and high runoff; 0.128 to 0.187 represents moderately vegetated areas with limited recharge potential; 0.187 to 0.257 includes dense vegetation, indicating better infiltration and moderate recharge conditions; 0.257 to 0.383 is linked to regions near streambeds or with very thick vegetation, supporting higher groundwater potential; and 0.383 to 0.748 suggests the presence of superficial water bodies and wetlands, which are key contributors to surface-subsurface water interactions. Under the ANN framework, the SAVI layer was assigned a contribution weight of 8% based on the model’s ability to identify nonlinear relationships between vegetation density and groundwater recharge. The ANN model highlights that areas with higher SAVI values generally correlate with increased recharge potential due to greater moisture availability and improved infiltration conditions. Conversely, lower SAVI zones were recognized as less conducive to groundwater recharge, aligning with their sparse vegetative cover and reduced moisture retention capacity. Thus, SAVI is a critical thematic input in the ANN-based delineation of GWPZs, offering valuable insights into the spatial distribution of recharge-friendly surfaces across the landscape.

NDWI

The NDWI is a crucial indicator in identifying and delineating water bodies across the landscape by utilizing the reflectance differences between green and near-infrared (NIR) wavelengths. Its sensitivity to moisture content and surface water presence makes it particularly effective for highlighting features such as rivers, lakes, reservoirs, and wetlands⁶³. However, NDWI can occasionally overestimate water presence, especially in areas affected by shadows, built-up surfaces, or saturated soils. NDWI assessments are often complemented with NDVI change detection analyses to enhance their reliability, which helps distinguish persistent water bodies from seasonal or misclassified features (Fig. 4(c)).

In this study, NDWI was computed to map spatial variations in water availability and surface saturation across the region. The NDWI values in the study area span from − 0.472 to 0.213, where higher values generally indicate greater surface water content or vegetation moisture, while lower values correspond to drier conditions or reduced moisture levels^64,65. Areas with higher NDWI values are typically associated with persistent water bodies or zones with saturated soil, which contribute positively to groundwater recharge, particularly in low-lying or flat terrains.

Within the ANN framework, NDWI plays a key role in characterizing surface moisture regimes and identifying potential zones for groundwater percolation. The ANN model evaluated NDWI alongside other input variables and determined its contribution to the groundwater potential to be 7%. This weight reflects the influence of surface water on recharge processes, especially in areas adjacent to rivers, canals, or reservoirs, where infiltration opportunities are enhanced. Integrating NDWI into the ANN-based model aids in capturing the hydrological interplay between surface water presence and subsurface recharge. By dynamically assessing spectral water signals, the model can account for direct (e.g., riverbank filtration) and indirect (e.g., saturated soils) mechanisms affecting GWPZs. This makes NDWI a vital component in the study area’s holistic analysis of recharge potential.

TGSI

The TGSI, though relatively less common in groundwater studies, is a valuable spectral indicator that captures subtle variations in surface texture and vegetation cover, which in turn influence infiltration potential and groundwater recharge. TGSI incorporates spatial information on soil moisture, surface grain size distribution, and vegetative cover - factors that collectively control runoff generation and percolation rates. Recent studies have demonstrated TGSI’s applicability in hydrological and infiltration-related assessments, validating its role in identifying recharge-prone areas^66,67.

In this study, TGSI values range from − 0.132 to 0.420, and the derived thematic map was classified into five hydrologically significant categories (Fig. 4(d)): −0.132 to 0.047 - Fine-grained soils (clay, silt) with high moisture retention, favorable for prolonged recharge. 0.047 to 0.073 - Coarser textures supporting moderate retention and infiltration. 0.073 to 0.088 - Grass-dominated surfaces with variable runoff potential depending on vegetation density and slope. 0.088 to 0.101 - Developed/urbanized zones with reduced infiltration due to impermeable surfaces. 0.101 to 0.420 - Bare or compacted surfaces with high runoff potential and minimal recharge capacity.

Field correlation analysis in the study area showed that low TGSI values were positively associated with higher infiltration rates and shallower groundwater depths, while higher TGSI values correlated with low recharge indicators. Within the ANN-based GWPZ model, TGSI was assigned a 5% contribution weight, reflecting a moderate yet meaningful influence on groundwater dynamics. Although certain indices such as NDVI and SAVI are known to be highly correlated due to their sensitivity to vegetation, each thematic layer in this study captures distinct environmental properties that influence groundwater recharge. TGSI’s inclusion alongside NDVI, SAVI, and NDWI enabled the model to capture surface textural, moisture, and vegetative transitions more effectively. Its contribution complements other parameters such as Slope, SAVI, and LULC, enhancing the model’s robustness and spatial accuracy in GWPZs mapping.

Slope

The slope is a critical topographic factor influencing groundwater potential, directly affecting surface runoff and water infiltration capacity. The steepness of the terrain determines how quickly precipitation flows over land, impacting the residence time available for infiltration^68,69. In the context of GWPZ mapping, slope data were derived from SRTM DEM using the slope tool in ArcGIS and categorized into five classes: 0–2% (Very gently sloping), 2–5% (Gently sloping), 5–15% (Moderat sloping), 15–35% (Steep loping), and > 35% (Very steep sloping) (Fig. 5(a)).In the ANN-based approach, very gently sloping regions (0–2%) emerged as the dominant influence class due to their higher potential for groundwater recharge. These areas allow maximum infiltration as water moves slowly across the surface, providing ample time for percolation into the subsurface. The ANN model identified a sharp decline in recharge potential as slope increases, especially in the > 15% categories. Steep and very steep slopes tend to accelerate runoff, significantly reducing the chance of water seepage and leading to minimal groundwater replenishment. The slope layer contributed 5% weight in the ANN model, indicating a moderate but indirect influence on groundwater potential. The ANN framework effectively recognized and learned from the spatial and statistical relationships between slope gradients and observed groundwater availability patterns, avoiding manual or subjective bias in weight assignment.

Fig. 5.

Thematic layers used in the study (continued). (a) Slope and (b) LULC. The slope layer was derived from the SRTM DEM, and the LULC layer was generated using Sentinel–2 satellite data. Both thematic layers were processed in ArcMap 10.8 (Esri, USA; https://www.esri.com/en-us/arcgis/products/arcgis-desktop/overview), with satellite datasets downloaded from the Copernicus Open Access Hub (https://scihub.copernicus.eu/). The final maps were overlaid on an OpenTopography basemap within ArcGIS to enhance terrain visualization and contextual interpretation.

Land Use/Land cover (LULC)

The LULC thematic map was generated through Supervised Classification, utilizing Sentinel-2 satellite imagery from 2022 with a spatial resolution of 10 m. Representative training samples − 70 per class - were carefully selected to enhance classification accuracy. The classification outcome was validated using a confusion matrix, with both overall accuracy and the Kappa coefficient confirming the map’s reliability and suitability for GWPZ delineation (Fig. 5(b)). The LULC map was classified into the following dominant land cover types: Waterbodies (822.68 km²), Forest (378.38 km²), Crops (18,221.06 km²), Built-up Area (2,141.11 km²), Barren Land (11,206.57 km²) These classes were further interpreted through the ANN-based framework. The ANN model identified cropland and barren land as dominant influence classes, as these areas typically allow better infiltration and groundwater recharge due to reduced surface sealing and minimal impervious cover. In contrast, built-up areas and dense forests, although ecologically valuable, exhibited lower influence in terms of recharge potential due to higher surface runoff and vegetation interception. This thematic layer was assigned a contribution weight of 10% in the ANN model, indicating its supporting yet significant role in shaping groundwater potential zones. By learning from the spatial distribution of land use types and correlating them with known recharge behaviors, the ANN system offered a nuanced understanding of LULC’s impact on subsurface hydrological dynamics.

GWPZS of MRB

The GWPZs map for the MRB was generated by integrating multiple thematic layers using a trained ANN model. The model learned the complex nonlinear relationships between input thematic parameters - such as geology, geomorphology, Slope, drainage density, lineament density, LULC, NDVI, NDWI, SAVI, and TGSI - and known groundwater conditions to predict groundwater potential more reliably than traditional heuristic methods. The final GWPZ map delineated five distinct zones based on the computed potential values: Very Good, Good, Moderate, Poor, and Very Poor (Fig. 6; Table 4). These zones were derived from the ANN-weighted overlay and spatial distribution analysis within ArcGIS. The ANN model provided data-driven contribution weights for each input layer, effectively capturing the combined influence of terrain, hydrological, geological, and vegetation-related factors.

Fig. 6.

Resultant GWPZs map of the MRB along with validation wells. The map was generated using ANN-based spatial modeling in ArcGIS 10.8 (https://www.esri.com/en-us/arcgis/products/arcgis-desktop/overview). The GWPZ output was overlaid on OpenStreetMap as the basemap within ArcGIS to enhance topographic and landmark visualization. Validation wells include both field-surveyed locations and Central Ground Water Board (CGWB) records, accessed from the CGWB Groundwater Data Portal (https://gwdata.cgwb.gov.in/).

Very Poor GWPZs (4,590.62 km² or 14.26%) occur primarily in regions with steep slopes, high runoff, low lineament and drainage densities, and unfavorable lithologies. These areas also display elevated TGSI values, high SAVI, and very low NDWI, all indicative of poor infiltration and recharge conditions. Poor zones (10,575.88 km² or 32.85%) exhibit slightly improved conditions but are, hindered by moderate slopes, less-fractured geology, and significant impermeable surfaces, limiting groundwater accumulation. Moderate zones (10,123.27 km² or 31.45%) are characterized by balanced hydro-geomorphological settings, with increased lineament presence, moderate vegetation (NDVI), and better infiltration potential - making them transitional zones for recharge. Good zones (5,541.25 km² or 17.21%) are associated with flat terrain, fractured crystalline rock formations, and vegetation conducive to infiltration. These regions appear prominently in the northeastern and southwestern parts of the MRB. Very Good GWPZs (1,360.54 km² or 4.23%) represent the most favorable recharge zones, supported by shallow slopes, high lineament, and waterbody density, optimal LULC types (croplands and barren areas), and low TGSI values. These areas form critical groundwater recharge hotspots.

Validation

ANN-based GWPZ map using AUC-ROC, confusion matrix, and well data

The correlation and multicollinearity analysis confirmed that all predictor variables were statistically independent, with correlation coefficients below 0.80 and VIF values less than 5. This ensures that no redundant information was introduced into the ANN framework, thereby enhancing model stability and generalization. Despite moderate correlation between NDVI and SAVI (r = 0.72), both were retained as they represent distinct vegetation and soil reflectance characteristics that influence groundwater recharge processes differently.

The performance of the ANN-derived GWPZ map was first validated using the AUC-ROC technique. The AUC value obtained was 0.850, indicating a high classification accuracy level and demonstrating the model’s strong predictive capability in distinguishing between different groundwater potential classes (Fig. 7(a)). The ROC curve illustrates the relationship between the true and false positive rates, providing a robust statistical measure for model validation. For this analysis, a total of 350 wells with measured groundwater levels were used, as AUC-ROC utilizes continuous probability scores for evaluation.

Fig. 7.

Validation of the ANN-derived GWPZs. (a) AUC–ROC curve showing model discrimination performance, (b) Spatial validation using well locations overlaid on the GWPZs map in ArcGIS 10.8 (https://www.esri.com/en-us/arcgis/products/arcgis-desktop/overview) with groundwater depth information obtained from field surveys and the CGWB Groundwater Data Portal (https://gwdata.cgwb.gov.in/). The wells show strong correspondence with the predicted groundwater potential classes, confirming the robustness of the ANN model.

In addition to AUC-ROC, a confusion matrix was generated to evaluate the multi-class classification performance of the model (Table 5). This analysis was based on 140 well locations where the depth to groundwater level could be clearly assigned to one of the five GWPZ categories. The matrix provides detailed insight into the correct and incorrect classifications for each category. From the confusion matrix, precision, recall, and F1-score were calculated for each class, along with macro-averaged values and the overall accuracy. The ANN model achieved an overall accuracy of 87%, with precision and recall values above 0.85 for most classes, indicating balanced predictive performance across all GWPZs (Table 6).

Table 5.

Confusion matrix for GWPZs classification using 140 Wells.

Actual \ predicted	Very good	Good	Moderate	Poor	Very poor	Total (actual)
Very good	14	1	0	0	0	15
Good	2	10	1	0	0	13
Moderate	0	3	35	3	0	41
Poor	0	0	5	46	1	52
Very poor	0	0	0	2	17	19
Total (predicted)	16	14	41	51	18	140

Table 6.

Classification metrics for GWPZs mapping (based on confusion matrix).

Class	Precision	Recall	F1-score
Very good	0.88	0.93	0.9
Good	0.71	0.77	0.74
Moderate	0.85	0.85	0.85
Poor	0.9	0.88	0.89
Very poor	0.94	0.89	0.91
Macro Avg.	0.86	0.86	0.86
Overall Accuracy	–	–	0.87

An overlay analysis of well location data and corresponding groundwater depth values was also conducted to verify the reliability of the ANN model. A total of 350 wells were used for this validation. The distribution pattern of wells across the five GWPZ categories aligned well with the model predictions. Specifically, 15 wells were recorded in the Very Good zones (depths: 15.24 to 19.81 m), 60 wells in Good zones (20.12 to 24.38 m), 110 wells in Moderate zones (24.69 to 33.53 m), 115 wells in Poor zones (33.83 to 41.15 m), and 50 wells in the Very Poor zones (depths > 141.15 m) (Fig. 7(b)).

The consistent spatial alignment between the ANN-derived GWPZ classification and observed field data, combined with strong AUC-ROC and multi-class performance metrics, confirms the robustness and accuracy of the ANN approach for delineating groundwater potential zones in the MRB. These results demonstrate that the model effectively integrates hydrogeological, environmental, and remote sensing parameters to identify areas with varying groundwater potential, thereby supporting sustainable groundwater resource management.

Spatial autocorrelation analysis of residuals

To assess the spatial independence of model prediction errors, a residual autocorrelation analysis was carried out using Moran’s I statistic⁷⁰. Residuals were calculated as the difference between the observed well-yield classifications and the corresponding ANN-predicted groundwater potential values. The analysis was conducted in ArcGIS 10.8, employing an inverse distance weighting (IDW) spatial relationship matrix to capture spatial proximity effects⁷¹. The computed Moran’s I value (0.037), along with a z-score of 0.74 and a p-value of 0.46, indicates that there is no statistically significant spatial autocorrelation among the residuals (Table 7). This finding implies that the prediction errors are spatially random rather than spatially clustered, confirming that the ANN model exhibits neither spatial bias nor localized overfitting within the MRB.

Table 7.

Moran’s i statistics for residual autocorrelation analysis.

Parameter	Moran’s I	Z-score	P-value	Interpretation
ANN-based GWPZ residuals	0.037	0.74	0.46	No significant spatial autocorrelation (random error distribution)

A residual semivariogram (Fig. 8) further substantiates this result. The semivariance values increase gradually with lag distance and stabilize beyond approximately 3 km, suggesting a weak or negligible spatial dependence structure in the residuals. This stabilization pattern demonstrates that prediction errors are randomly distributed across the study area, consistent with the Moran’s I outcome. Together, these results validate the spatial generalization capability and robustness of the ANN-based GWPZ model, ensuring reliable GWPZs mapping across heterogeneous terrains in the MRB.

Fig. 8.

Residual semivariogram for ANN model.

Inferential statistical validation using Kruskal - Wallis test

To statistically validate the differentiation among ANN-predicted groundwater potential classes, a Kruskal - Wallis H test was performed between observed groundwater depths (m bgl) and their corresponding GWPZ categories^72–74. This non-parametric test was selected because groundwater depth data did not meet the assumption of normality. The analysis was conducted using SPSS software at a 95% confidence level (α = 0.05).

The Kruskal - Wallis test revealed a statistically significant difference in groundwater depths among the five GWPZ classes (H = 42.87, df = 4, p < 0.001), confirming that the ANN-based zonation effectively distinguishes hydrogeological variability across the study area. Post-hoc pairwise comparisons using Dunn’s test with Bonferroni correction indicated that the Very Good and Good zones have significantly shallower groundwater levels compared to the Poor and Very Poor zones, while the Moderate class represented an intermediate condition (Tables 8 and 9).

Table 8.

Kruskal - Wallis test results for observed groundwater depths vs. Predicted GWPZ Classes.

GWPZ class	Mean rank	Median depth (m bgl)	No. of wells
Very good	85.3	17.8	15
Good	118.7	21.6	60
Moderate	179.2	28.7	110
Poor	231.5	33.1	115
Very poor	294.6	41.8	50
Test statistic (H)	42.87
Degrees of freedom (df)	4
p-value	< 0.001

Table 9.

Summary of well data and depth distribution across GWPZ Classes.

GWPZ class	No. of wells	Mean depth (m bgl)	Median depth (m bgl)	Min depth (m bgl)	Max depth (m bgl)
Very good	15	17.2	17.8	15	19.5
Good	60	21.3	21.6	19	25
Moderate	110	28.1	28.7	26	32
Poor	115	32.9	33.1	30	36.5
Very poor	50	41.2	41.8	39	45.5

These results demonstrate that the ANN-derived GWPZ map accurately reflects real groundwater depth variations, thereby validating the model’s predictive robustness and practical applicability in hydrogeological zoning and groundwater resource planning.

The progressive increase in both mean rank and median depth across the five groundwater potential classes confirms that groundwater depth systematically varies with the ANN-predicted zones. This strong statistical consistency highlights the model’s hydrogeological validity and decision-making utility. Figure 9 illustrates the observed variation in groundwater depth across the ANN-predicted classes, showing a clear upward trend from Very Good to Very Poor zones consistent with actual field conditions.

Fig. 9.

Variation in groundwater depth across the ANN-predicted classes.

Feature importance and sensitivity analysis

To evaluate the relative contribution of each thematic layer in predicting groundwater potential, a feature sensitivity (drop-one-feature) analysis was performed. In this approach, each predictor variable was sequentially removed from the input dataset, and the ANN model was retrained to assess changes in accuracy and AUC. The results revealed that exclusion of NDVI, SAVI, or TGSI led to a performance decline of 3.5–5.2%, indicating their significant role in capturing vegetation and surface texture characteristics. Other parameters such as slope, drainage density, and LULC showed moderate influence, whereas soil-related attributes exhibited relatively stable responses. The dataset was consistently partitioned into 70% training and 30% testing subsets with a fixed random seed to ensure reproducibility. These results confirm that each predictor contributes unique and complementary information to the ANN framework, enhancing the robustness and interpretability of the GWPZ classification.

Uncertainty and future work

In the current study, model uncertainty has been assessed indirectly through quantitative validation metrics, including AUC-ROC, confusion matrix accuracy, class-specific performance indices, Moran’s I residual spatial autocorrelation, and the Kruskal–Wallis statistical test. While these metrics collectively demonstrate strong predictive reliability, spatially explicit uncertainty products, such as pixel-level residual/error maps, bootstrap-derived confidence surfaces, or ensemble spread maps were not generated. Consequently, the present version of the ANN-based GWPZ model does not provide pixel-scale confidence information, which limits the direct communication of spatial prediction certainty to groundwater planners and decision-makers.

Future work will focus on incorporating spatial uncertainty modelling by generating: (i) pixel-wise residual maps derived from independent validation wells, (ii) bootstrapped ensemble confidence surfaces quantifying prediction variability, and (iii) spatial error propagation assessments across thematic layer inputs. The integration of such uncertainty products will strengthen the interpretability and utility of GWPZ outputs, support risk-aware groundwater exploration, and improve the transparency of recommendations for artificial recharge and well-site prioritization.

Discussion

The integration of ANN in delineating GWPZs across the MRB demonstrated a significant advancement over traditional weighted models like the Analytical Hierarchy Process (AHP). The ANN model, trained using multiple thematic layers and validated through field well data, effectively captured nonlinear relationships among diverse hydrogeological and land surface variables. This approach enhanced the classification accuracy and spatial precision of the resultant GWPZ map, as evidenced by an AUC-ROC value of 0.850, indicating high model sensitivity and predictive performance.

The network architecture was designed with two hidden layers (32 and 16 neurons) and five output neurons with a Softmax activation function, directly corresponding to the five GWPZ classes viz., Very Good, Good, Moderate, Poor, and Very Poor, ensuring methodological validity in multi-class classification. Similar ANN-based groundwater potential studies, have adopted comparable architectures (one or two hidden layers, 10–30 neurons) and achieved accuracies between 85% and 94%, which aligns with the present study’s performance under semi-arid conditions^75,76.

Spatial analysis indicated that slope, geology, lineament density, LULC, SAVI, NDVI, NDWI, and TGSI were key determinants of groundwater occurrence and recharge potential. Prior to ANN modeling, multicollinearity diagnostics (correlation and VIF analysis) confirmed the statistical independence of input predictors, ensuring that each thematic factor contributed unique information to the model’s learning process. This strengthened the reliability of the derived contribution weights and minimized bias arising from correlated inputs, thereby improving the interpretability of the ANN-based predictions. Areas with gentle slopes, fractured crystalline formations, high lineament density, and extensive croplands demonstrated higher groundwater potential. In contrast, regions with steep slopes, sparse vegetation, rapid urban expansion, and silty substrates showed limited infiltration capacity. These results are consistent with earlier findings, which reported that machine learning models excel in heterogeneous terrains by effectively capturing nonlinear relationships and interdependencies among variables^77,78.

The overlay of well data and groundwater depth corroborated the ANN-classified zones, with shallow water tables predominantly occurring in “Very Good” and “Good” zones. This consistency between predicted potential and empirical measurements reinforces the model’s validity and reliability. Nevertheless, the accuracy of ANN-based outputs remains contingent upon the input datasets’ spatial resolution, completeness, and currency. Not fully accounted for in the model, seasonal variations and anthropogenic influences on groundwater dynamics may introduce marginal uncertainties. Despite these constraints, the ANN-based GWPZ mapping offers a scalable and adaptive regional water resource planning framework. The resultant maps are significant for decision-makers in prioritizing artificial recharge sites, optimizing groundwater extraction, and fostering sustainable watershed development. Moreover, this methodology can be transferred to other data-scarce abasins where conventional models fail to capture spatial heterogeneity and complexity in groundwater systems.

Conclusion

The escalating demand and unsustainable extraction of groundwater resources in arid and semi-arid regions, such as the MRB, necessitate advanced and reliable methods for assessing groundwater potential. This study adopted an ANN-based approach integrated with geospatial tools to delineate GWPZs, offering a data-driven alternative to traditional methods like the AHP. The model incorporated ten thematic layers, including geology, geomorphology, slope, drainage and lineament densities, LULC, NDVI, TGSI, NDWI, and SAVI, each playing a critical role in influencing groundwater dynamics. The ANN-based model effectively captured these variables’ complex, nonlinear interactions, resulting in a more robust and spatially accurate GWPZ map.The results revealed that approximately 62.08% of the MRB falls within moderate to very poor GWPZs, while only 21.44% lies in good to very good zones, underscoring the urgency for sustainable groundwater management. The validation using AUC-ROC (0.850) and well data overlay confirmed the predictive reliability of the ANN model, affirming its suitability for hydrological zoning in complex geological terrains. However, the study has some limitations. The absence of high-resolution subsurface geological and aquifer data restricted the model’s capacity to fully represent deeper hydrogeological processes. The seasonal fluctuations in groundwater levels and anthropogenic influences such as land-use change and groundwater abstraction trends were not dynamically modeled. Future research should integrate geophysical, hydrogeological, and time-series hydrometeorological datasets to refine groundwater potential assessments and enable spatiotemporal analysis. Despite these limitations, the study contributes significantly to hydroinformatics and spatial decision support systems. The generated GWPZ maps are vital tools for groundwater exploration, well-site selection, and recharge planning. For policymakers and planners, these outputs offer scientific guidance in promoting equitable and sustainable groundwater extraction, especially in vulnerable agricultural and rural regions. This research directly aligns with the United Nations Sustainable Development Goals (SDGs), particularly Goal 6 (Clean Water and Sanitation) and Goal 13 (Climate Action), by fostering informed water management in semi-arid regions. The integration of ANN with remote sensing and GIS enhances the precision of GWPZ mapping by effectively handling complex spatial datasets and nonlinear relationships among hydrogeological factors. Moreover, the use of statistical validation methods, such as AUC-ROC and confusion matrices, reinforces the robustness and reliability of the model predictions, offering confidence for practical applications. While uncertainties inherent to input data and model assumptions remain, the study demonstrates that ANN-based approaches provide a scalable and adaptable framework for groundwater assessment, contributing meaningfully to both academic research and sustainable water resource management.

Author contributions

Dr. Badapalli Pradeep Kumar: Manuscript writing, methodology, Visulaization. Mrs. Anusha Boya Nakkala: Manuscript writing, Visulaization. Dr. Padma Sree Pujari: Writing, review and editing. Dr. Sakram Gugulothu: Supervision, Writing, review and editing. Prof.Raghu Babu Kottala: Validation, data collection. Dr. Prasad Mannala: Writing, review and editing, and Data collection. Wafa Saleh Alkhuraiji: Writing, review and editing, Resources. Kgabo Humphrey Thamaga: Writing, review and editing. Mohamed Zhran: Investigation, Writing, review and editing, Resources, Project administartion. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2026R680), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Data availability

Data will be made available at the request of the Corresponding author.

Declarations

Competing interests

The authors declare no competing interests.

Consent for participate

Not applicable.

Consent for publication

Not applicable.

Ethical approval

Not applicable.