Assessing and Forecasting Groundwater Resources in the Context of Climate Change Using AI Techniques for the Industry Zones in Tiruppur, India

Abstract

Groundwater systems in semi-arid and industrial regions are increasingly affected by climate-driven non-stationarity and anthropogenic pressure, challenging conventional forecasting approaches. This study develops and evaluates an integrated artificial intelligence framework designed to minimize piezometric head residual dispersion under non-stationary hydroclimatic conditions. The proposed methodology combines Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) and Variational Mode Decomposition (VMD) with a Slime Mould Algorithm–optimized Long Short-Term Memory (SMA–LSTM) model and a CNN–LSTM architecture, which are dynamically fused using an Adaptive Weighting Model (AWM). The framework was applied to long-term groundwater level (1994–2024), groundwater quality (2017–2023), and meteorological datasets to evaluate the predictive robustness across climatic variability regimes. The proposed ensemble achieved a mean absolute error of 0.267 m, root mean square error of 0.429 m, coefficient of determination (R²) of 0.948, and Nash–Sutcliffe efficiency of 0.938, representing substantial residual reduction compared to baseline deep learning models. Residual diagnostics confirmed minimized peak deviations and stable performance under non-stationary conditions. Scenario-based simulations driven by CMIP6 climate projections indicate increasing groundwater stress under future warming trajectories, with amplified variability and declining recharge signals. These findings demonstrate that multi-stage signal decomposition coupled with metaheuristic optimization and adaptive ensemble learning significantly enhances predictive stability and residual minimization in climate-sensitive aquifer systems. The proposed framework provides a transferable, climate-resilient decision-support tool for sustainable groundwater management in industrial and semi-arid regions.

1. Introduction

Groundwater constitutes one of the most critical freshwater resources worldwide, supporting domestic consumption, agricultural irrigation, and industrial production. In semi-arid and climate-sensitive regions, groundwater frequently represents the primary and most reliable water source due to limited and highly variable surface water availability [1]. Rapid population growth, accelerated urbanization, and expanding industrial activity have intensified groundwater abstraction across many developing regions, resulting in persistent declines in water tables and the progressive deterioration of groundwater quality. These pressures are particularly severe in hard-rock aquifer systems, where storage capacity is inherently constrained and recharge processes are strongly dependent on episodic rainfall events [2,3].

Climate change further compounds groundwater stress by altering precipitation regimes, increasing evapotranspiration rates, and intensifying the frequency and magnitude of extreme hydrological events [2]. These hydroclimatic alterations disrupt recharge dynamics and amplify the nonlinear and non-stationary behavior of groundwater systems. Recent investigations across semi-arid regions, including the Indian peninsula, have demonstrated strong correlations between groundwater level fluctuations and large-scale climate variability indices, emphasizing the necessity of incorporating climatic signals into groundwater assessment and prediction frameworks [4]. The increasing variability in recharge patterns and abstraction pressures therefore necessitates forecasting approaches capable of capturing multi-scale temporal dependencies and climate-induced non-stationarity.

Traditional groundwater forecasting methods, including statistical regression techniques and physically based numerical models, often struggle to adequately represent the complex interactions between anthropogenic withdrawals, climatic variability, and heterogeneous aquifer properties [5,6]. In response to these limitations, machine learning and deep learning techniques have gained substantial attention for groundwater level prediction. Among these, Long Short-Term Memory networks have demonstrated strong capability in modeling temporal dependencies within groundwater time series. However, standalone LSTM architectures may exhibit reduced robustness when applied to highly non-stationary systems influenced by climate variability and abrupt recharge shifts.

To enhance predictive stability, recent research has explored hybrid modeling strategies that integrate signal decomposition techniques with deep learning architectures. Approaches combining empirical mode decomposition, wavelet transforms, and recurrent neural networks have shown improved forecasting performance by isolating intrinsic temporal components prior to model training [1,7]. Ensemble deep learning frameworks and multi-modal decomposition strategies have further improved robustness by leveraging complementary feature representations. Simultaneously, advances in metaheuristic optimization algorithms have been employed to fine-tune model hyperparameters and improve convergence behavior under complex hydroclimatic conditions. Emerging architectures such as hybrid temporal convolutional network–LSTM models with attention mechanisms have also demonstrated enhanced capability in extracting spatial–temporal features from groundwater datasets.

Despite these methodological advancements, existing studies frequently address signal non-stationarity, deep learning optimization, or ensemble integration in isolation, without systematically combining these components into a unified and climate-resilient forecasting framework. Moreover, comparatively limited emphasis has been placed on explicitly minimizing piezometric head residual dispersion under climate-driven non-stationary conditions, which remains a critical requirement for reliable groundwater management and decision-making. Accordingly, this study posits that integrating multi-stage signal decomposition with metaheuristic-optimized deep learning models and adaptive ensemble weighting can significantly enhance the groundwater level forecasting accuracy and reduce residual variability relative to standalone deep learning approaches in climate-influenced aquifer systems.

To evaluate this premise, the present work develops an integrated forecasting framework that combines ICEEMDAN- and VMD-based signal decomposition with a Slime Mould Algorithm–optimized LSTM model and a CNN–LSTM architecture, which are dynamically fused using an Adaptive Weighting Model. The framework is assessed using long-term groundwater level, groundwater quality, and meteorological datasets, and its robustness is evaluated under historical climate variability and CMIP6-based future climate projections. Through this integrated approach, the study aims to advance climate-resilient groundwater forecasting methodologies applicable to semi-arid and industrially stressed aquifer systems.

Recent advances in machine learning have increasingly incorporated metaheuristic optimization algorithms for feature selection and hyperparameter tuning. Metaheuristic approaches provide effective solutions for complex nonlinear optimization problems where traditional gradient-based methods may struggle to converge. A comprehensive survey by [8] highlights the growing application of algorithms such as genetic algorithms, particle swarm optimization, and slime mould algorithms in improving model accuracy and computational efficiency. These approaches are particularly valuable in environmental and hydrological modeling where datasets are highly nonlinear and non-stationary. Metaheuristic-driven machine learning models have also been successfully applied in groundwater systems. For example, recent studies demonstrate that integrating genetic and iterative metaheuristic algorithms with predictive models can significantly enhance the groundwater modeling accuracy and drought inference capabilities [9]. These hybrid optimization frameworks improve parameter calibration and enable more reliable predictions in complex hydrogeological systems. Machine learning techniques have been increasingly adopted for groundwater level prediction due to their ability to capture nonlinear relationships between hydroclimatic variables. Previous studies have demonstrated that machine learning models combined with geostatistical approaches can significantly improve the groundwater level prediction accuracy and spatial interpolation of groundwater variability [6]. Recent studies have also emphasized the need for integrated frameworks that simultaneously address non-stationarity, model optimization, and ensemble learning in groundwater systems [1,7,10]. However, such comprehensive approaches remain limited, particularly under climate-driven variability, motivating the development of the proposed framework.

To address the limitations of existing groundwater forecasting approaches under climate-induced non-stationarity, the primary objective of this study was to develop a robust and integrated artificial intelligence-based framework for accurate groundwater level prediction. This study focused on combining multi-stage signal decomposition techniques, including ICEEMDAN and Variational Mode Decomposition (VMD), to effectively handle non-stationary groundwater time series. Furthermore, deep learning models such as LSTM and CNN–LSTM were developed and optimized using the Slime Mould Algorithm (SMA) to enhance predictive performance and convergence efficiency. To improve forecasting robustness and minimize residual errors, the outputs of individual models were integrated through an Adaptive Weighting Model (AWM). Finally, the proposed framework was evaluated under both historical conditions and future climate scenarios using CMIP6 projections to assess its applicability in climate-sensitive environments. Through this integrated approach, the study aims to advance climate-resilient groundwater forecasting methodologies for semi-arid and industrially stressed regions.

In addition to technical approaches, recent studies have emphasized the importance of institutional and policy frameworks in groundwater management. Liptrot and Hussein [11] demonstrated that regulatory interventions, including tariffs and licensing, often fail to significantly reduce groundwater abstraction due to enforcement limitations and socio-political resistance. In addition to technical approaches, recent studies have emphasized the importance of institutional and policy frameworks in groundwater management.

2. Materials and Methods

2.1. Study Area

The study area comprises Tiruppur District, situated in the western part of Tamil Nadu, India, between approximately 11.0–11.5° N latitude and 77.0–77.6° E longitude. The district lies within a semi-arid climatic zone, characterized by high inter-annual rainfall variability, frequent dry spells, and increasing temperature trends. Average annual rainfall is primarily governed by the southwest and northeast monsoons, with recharge largely concentrated during short, intense precipitation events.

Hydro-geologically, Tiruppur is dominated by hard-rock formations, mainly consisting of weathered and fractured crystalline rocks. Groundwater occurrence in such terrains is highly heterogeneous and is largely confined to shallow weathered zones and deeper fracture networks. The limited storage capacity of these aquifers, coupled with low natural recharge rates, makes groundwater availability extremely sensitive to variations in rainfall intensity, duration, and seasonal distribution. Consequently, groundwater levels exhibit pronounced temporal fluctuations and spatial variability across the district.

Tiruppur is internationally recognized as the “Knitwear Capital of India”, hosting a dense concentration of textile dyeing and processing industries. Rapid industrialization, combined with urban expansion and agricultural water demand, has led to the persistent over-extraction of groundwater for industrial, domestic, and irrigation purposes. In the absence of perennial surface water sources, groundwater remains the primary source of water supply, further intensifying the stress on aquifer systems.

In addition to quantitative depletion, groundwater quality deterioration has emerged as a critical concern. Discharge of untreated or partially treated textile effluents, rich in dissolved salts and chemical residues, has resulted in elevated levels of total dissolved solids (TDS), electrical conductivity, and salinity in several parts of the district. These impacts are exacerbated by reduced dilution during low-recharge periods, particularly under prolonged drought conditions.

Climate change further compounds groundwater vulnerability in the region by altering monsoon patterns, increasing evapotranspiration, and amplifying the frequency of extreme hydrological events. These combined hydrogeological, industrial, and climatic factors make Tiruppur District a representative and high-risk case study for evaluating advanced, data-driven groundwater forecasting methodologies. The complex and non-stationary nature of groundwater dynamics in this region provides an ideal testbed for assessing the effectiveness of hybrid deep learning models under climate-induced stress conditions.

Consequently, groundwater circulation in the region exhibits strong spatial heterogeneity and is highly sensitive to rainfall variability, fracture connectivity, and localized recharge conditions, particularly under semi-arid climatic settings. Similar hydrogeological behavior has been reported in other complex fractured aquifer systems, where stochastic groundwater dynamics are influenced by both climatic variability and anthropogenic abstraction pressures [3,12]. This inherent complexity motivates the use of advanced data-driven modeling approaches capable of capturing nonlinear and non-stationary groundwater responses.

The Figure 1 shows the geographic location of Tiruppur District within Tamil Nadu and India using inset maps, along with the spatial distribution of groundwater monitoring wells across Tiruppur District.

The figure illustrates the conceptual hydrogeological cross-section of the study area, showing the shallow weathered zone, underlying fractured rock aquifer, and deeper hard-rock basement. Groundwater occurrence is primarily confined to the weathered and fractured zones, with recharge occurring through rainfall infiltration and flow governed by fracture connectivity. To provide a clearer representation of the subsurface aquifer configuration, a conceptual hydrogeological cross-section of the study area is presented in Figure 2. The cross-section illustrates the typical hard-rock aquifer system of Tiruppur District, characterized by a shallow weathered zone overlying fractured crystalline formations and a deeper low-permeability basement. Groundwater occurrence is primarily confined to the weathered mantle and fracture networks, with recharge driven by monsoonal rainfall and flow controlled by fracture connectivity and spatial heterogeneity.

2.2. Data Collection and Sources

Recent advancements in monitoring systems have enabled the collection of high-resolution groundwater data suitable for data-driven modeling. IoT-based and telemetry-enabled groundwater observation networks have been successfully utilized to support machine learning-based groundwater level prediction [13]. In this study, groundwater level and quality data from both manual and telemetry-based monitoring wells operated by central and state agencies were integrated with meteorological observations to ensure comprehensive spatial and temporal coverage. Groundwater observations from multiple monitoring wells were aggregated to represent regional groundwater dynamics across Tiruppur District. This aggregation approach facilitates the development of a generalized forecasting framework; however, it may smooth localized variations associated with spatial heterogeneity in aquifer characteristics and abstraction patterns.

2.2.1. Groundwater Level Data

Groundwater level data were obtained from the Water Resources Department (WRD), Government of Tamil Nadu, covering a continuous period from 1994 to 2024. The dataset consists of monthly observations of depth to groundwater level (meters below ground level) collected from a network of observation wells distributed across Tiruppur District. Each record is accompanied by spatial attributes, including well identification number, geographic coordinates (latitude and longitude), administrative boundaries (taluk and village), and well type (dug wells and bore wells).

This long-term dataset captures seasonal recharge patterns, inter-annual variability, and long-term depletion trends driven by industrial abstraction and land-use change. The extended temporal coverage makes the dataset particularly suitable for deep learning-based time-series modeling and for assessing non-stationary groundwater behavior under evolving climatic conditions.

2.2.2. Groundwater Quality Data

Groundwater quality data were collected from the Tamil Nadu Water Supply and Drainage (TWAD, Government of Tamil Nadu, Chennai, India) Board and the Central Ground Water Board (CGWB, Ministry of Jal Shakti, New Delhi, India) for the period covering 2017–2023. The dataset includes key physicochemical parameters that indicate groundwater suitability for domestic and agricultural use, such as pH, electrical conductivity (EC), and total dissolved solids (TDS). In addition, major ionic constituents—including calcium (Ca), magnesium (Mg), sodium (Na), chloride (Cl), sulfate (SO), and bicarbonate (HCO)—were considered.

Derived indices such as sodium adsorption ratio (SAR) and residual sodium carbonate (RSC) were computed to assess salinity hazards and irrigation suitability. These parameters provide critical insight into groundwater quality degradation resulting from industrial effluent discharge, excessive abstraction, and reduced natural dilution during low-recharge periods. The inclusion of groundwater quality data allows the proposed framework to evaluate both the quantitative and qualitative aspects of groundwater sustainability. Although nitrate (NO₃⁻) is an important indicator of anthropogenic contamination, particularly from agricultural activities and wastewater sources, it was not included in the present study due to the lack of consistent and continuous data across all monitoring locations for the selected time period. The inclusion of parameters with incomplete spatial or temporal coverage could introduce bias and affect the robustness of the modeling framework. Instead, emphasis was placed on key hydrochemical parameters such as electrical conductivity, total dissolved solids, and major ions, which provide a reliable representation of groundwater salinity, mineralization, and industrial contamination—dominant factors influencing groundwater quality in the study area. Groundwater quality parameters were incorporated primarily for exploratory analysis and to support the understanding of hydrochemical variability in the study area. Due to the relatively shorter temporal coverage (2017–2023), these parameters were not directly used as primary targets in the forecasting framework but were utilized to analyze relationships with groundwater levels and to inform feature engineering.

2.2.3. Meteorological and Climate Data

Meteorological data were obtained from the Indian Meteorological Department (IMD) and included daily observations of rainfall (mm), maximum and minimum temperature (°C), and relative humidity (%) for the period 1994–2024. These variables are key climatic drivers influencing groundwater recharge, evapotranspiration, and seasonal water-level fluctuations. Daily records were aggregated to a monthly temporal resolution to ensure consistency with the groundwater observations.

To assess future groundwater behavior under climate change, climate projection data from the Coupled Model Intercomparison Project Phase 6 (CMIP6 via WorldClim, Fick and Hijmans, University of California, Davis, CA, USA) were incorporated. Projections were considered under multiple Shared Socioeconomic Pathways (SSPs), including the SSP2-4.5 (moderate emissions) and SSP5-8.5 (high emissions) scenarios. These datasets provide downscaled estimates of future precipitation and temperature patterns and enable long-term groundwater forecasting under alternative climate trajectories.

2.2.4. Data Integration and Temporal Alignment

All datasets were harmonized by resampling to a common monthly temporal resolution and aligned spatially using geographic coordinates. This integration facilitates the seamless fusion of groundwater, quality, and climatic variables within the modeling framework.

A summary of the datasets, sources, temporal coverage, and parameters is presented in

Table 1. Summary of datasets used in this study.

Dataset	Source	Time Span	Parameters
Groundwater Level	WRD, Tamil Nadu 1	1994–2024	Depth to Water Level, Location ID, Latitude, Longitude
Groundwater Quality	TWAD, CGWB 2	2017–2023	pH, EC, TDS, Ca, Mg, Na, Cl, SO4
Meteorological Data	IMD 3	1994–2024	Rainfall, Temperature, Humidity
Climate Projections	CMIP6 4	2021–2040	Precipitation, Temperature under Medium Emission (SSP2-4.5), and Very High Emission (SSP5-8.5)

¹ Groundwater Level Dataset—WRD, Tamilnadu—https://www.tn.gov.in/dept_profile.php?dep_id=NDQ= (accessed on 23 March 2025) ² Groundwater Quality Dataset—TWAD, CGWB—https://cgwb.gov.in/en (accessed on 20 March 2025) ³ Metrological Data—IMD—https://mausam.imd.gov.in/ (accessed on 21 March 2025) ⁴ Climate Projections—CMIP6—https://www.worldclim.org/data/index.html (accessed on 22 March 2025).

.

The Figure 3 depicts the overall architecture of the proposed hybrid artificial intelligence-based groundwater forecasting framework. The framework integrates historical groundwater level and quality data with meteorological inputs, applies ICEEMDAN–VMD-based signal decomposition and feature engineering, and employs SMA-optimized LSTM and CNN–LSTM models whose outputs are combined using an adaptive weighting model to generate the final groundwater predictions, which are evaluated using standard performance metrics.

2.2.5. Data Accuracy and Quality Assurance

Ensuring the accuracy and reliability of the input data is critical for developing robust groundwater forecasting models. The datasets used in this study were obtained from established government agencies, including the Water Resources Department (WRD, Government of Tamil Nadu, Chennai, India), the Tamil Nadu Water Supply and Drainage (TWAD, Government of Tamil Nadu, Chennai, India) Board, the Central Ground Water Board (CGWB, Ministry of Jal Shakti, New Delhi, India), and the Indian Meteorological Department (IMD, New Delhi, India), all of which follow standardized protocols for data collection and quality control.

Groundwater level measurements were recorded using calibrated observation wells with consistent monitoring procedures, ensuring high temporal reliability. Groundwater quality parameters were obtained from laboratory analyses conducted under standard hydrochemical testing guidelines, ensuring acceptable measurement precision and reproducibility. Meteorological data from IMD are subject to rigorous validation and quality assurance procedures, including instrument calibration and cross-verification.

To further enhance data reliability, preprocessing steps such as missing value imputation, outlier detection using statistical methods (Z-score and interquartile range), and consistency checks were applied. These procedures minimize the impact of measurement errors and ensure that the datasets used for modeling accurately represent the underlying hydroclimatic conditions. Consequently, the integrated dataset was considered reliable and suitable for advanced machine learning-based groundwater forecasting.

2.3. Data Preprocessing and Cleaning

Data preprocessing is a critical step in developing reliable machine learning and deep learning models, particularly when working with heterogeneous hydroclimatic datasets characterized by missing values, measurement noise, and temporal inconsistencies. In this study, a comprehensive preprocessing pipeline was implemented to ensure data quality, consistency, and suitability for advanced time-series modeling. These preprocessing steps also serve as a secondary validation mechanism to ensure data quality and reliability prior to model training.

2.3.1. Handling Missing and Invalid Data

Missing values in groundwater datasets were handled using linear interpolation and forward-filling to preserve temporal continuity. Invalid or non-numeric entries were removed based on data quality criteria.

2.3.2. Temporal Resampling and Alignment

All datasets were resampled to a monthly temporal resolution to ensure consistency. Meteorological variables were aggregated accordingly, and temporal alignment was performed to synchronize all input features.

2.3.3. Unit Standardization and Data Transformation

All variables were standardized to consistent units, and logarithmic transformation was applied to selected skewed parameters to improve model stability.

2.3.4. Normalization and Scaling

Feature scaling was performed using min–max normalization to map all variables into the range [0, 1], improving model convergence.

2.3.5. Outlier Detection and Treatment

Outliers were identified using Z-score and interquartile range methods and treated appropriately to reduce noise while preserving significant hydrological variations.

2.3.6. Data Consistency and Quality Assurance

Final datasets were validated through consistency checks to ensure reliability and suitability for time-series modeling.

2.4. Feature Engineering

Feature engineering was performed to enhance model performance by capturing the temporal, statistical, and climatic characteristics of groundwater dynamics. In groundwater forecasting, appropriately designed features enable models to learn seasonal behavior, long-term trends, and delayed system responses driven by climatic and anthropogenic factors. In this study, a systematic feature engineering strategy was adopted to improve model interpretability and forecasting accuracy.

After feature construction and selection, a total of 21 representative features were retained for model training, capturing the dominant temporal, statistical, and seasonal characteristics of groundwater dynamics while avoiding redundancy.

2.4.1. Time-Based Features

Time-based features such as month, season, and long-term trends were included to capture temporal variability in groundwater levels.

2.4.2. Lagged Features

Lagged variables were incorporated to account for delayed groundwater responses to climatic inputs.

2.4.3. Rolling Statistical Features

Rolling statistical features were used to capture short-term variability and long-term trends in groundwater behavior.

2.4.4. Climate Interaction Features

Interaction features combining rainfall, temperature, and humidity were generated to capture nonlinear climate–groundwater relationships.

2.4.5. Feature Selection and Dimensionality Control

Feature selection was performed using correlation analysis to remove redundant variables and retain relevant predictors. To mitigate the risk of overfitting due to a high feature-to-sample ratio, a threshold-based correlation filtering approach was applied. Features exhibiting a Pearson correlation coefficient greater than 0.85 were considered redundant, and only one representative feature from each correlated group was retained. This ensured dimensionality reduction while preserving informative variables.

In addition, feature stability was assessed by evaluating the correlation consistency across different temporal subsets of the dataset. Only features demonstrating stable relationships with the target variable were retained for modeling. This approach improves robustness and reduces the likelihood of spurious correlations influencing model performance.

2.5. Signal Decomposition

Groundwater time series are inherently non-stationary due to the combined influence of seasonal recharge, long-term abstraction trends, and climate variability. Signal decomposition techniques provide an effective means of isolating these overlapping temporal components prior to model training. Previous studies have demonstrated that decomposition-assisted deep learning frameworks outperform standalone models in groundwater forecasting tasks [7].

In this study, the ICEEMDAN decomposition was configured to extract a maximum of 8 intrinsic mode functions (IMFs), based on the energy distribution and variance contribution of the signal. The decomposition process was terminated when the residual component became a monotonic function or when the standard deviation between successive iterations fell below a predefined threshold (ε = 0.2). High-frequency IMFs (typically IMF1–IMF3), representing short-term fluctuations and noise, were selected for secondary decomposition using Variational Mode Decomposition (VMD). The number of VMD modes (K) was set to 3, and the penalty parameter (α) was fixed at 2000 to balance bandwidth constraint and reconstruction accuracy. This two-stage decomposition ensures the effective separation of noise-dominated and trend-dominated components.

The Figure 4 illustrates the signal decomposition and preprocessing workflow for groundwater and meteorological time-series data. Historical groundwater observations were decomposed using Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) to obtain intrinsic mode functions, which were further refined using Variational Mode Decomposition (VMD). The resulting sub-signals and preprocessed meteorological inputs were used as inputs for subsequent feature engineering and model development.

2.5.1. Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN)

Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) is an enhanced variant of empirical mode decomposition designed to overcome mode mixing and residual noise commonly observed in conventional EMD approaches. ICEEMDAN decomposes a complex time series into a finite set of Intrinsic Mode Functions (IMFs) and a residual trend component, each representing oscillations at distinct temporal scales.

In this study, ICEEMDAN was applied to groundwater level and quality time series to extract meaningful components corresponding to high-frequency fluctuations, seasonal variations, and long-term trends. By introducing adaptive noise during decomposition, ICEEMDAN ensures improved stability and consistency across ensembles, resulting in clearer separation of hydrological signals influenced by climatic and anthropogenic factors.

2.5.2. Variational Mode Decomposition

While ICEEMDAN effectively separates intrinsic components, certain high-frequency IMFs may still contain residual noise. To further refine these components, Variational Mode Decomposition (VMD) was employed as a secondary decomposition step. VMD formulates signal decomposition as a constrained variational problem, extracting modes with specific bandwidths and center frequencies.

The application of VMD to selected ICEEMDAN-derived IMFs enhances frequency resolution and suppresses spurious oscillations. This two-stage decomposition strategy improves the clarity of temporal features supplied to deep learning models, enabling more accurate learning of groundwater dynamics across multiple time scales.

2.5.3. Integration with Deep Learning Models

Each decomposed component obtained through ICEEMDAN and VMD was used as an independent input to deep learning models, particularly LSTM-based architectures. By training models on simplified and scale-specific signals rather than raw time series, the forecasting framework reduces model complexity and mitigates overfitting. The final groundwater prediction is reconstructed by aggregating outputs from individual decomposed components.

This integrated decomposition–learning approach enables the model to capture both short-term variability and long-term groundwater trends more effectively than conventional single-stage models. The decomposition process thus plays a critical role in enhancing prediction accuracy, stability, and interpretability under climate-induced non-stationary conditions.

2.6. Model Development

Advanced deep learning architectures have been increasingly adopted for groundwater level prediction due to their ability to capture complex temporal and spatial dependencies. Recent studies employing hybrid architectures, such as TCN–LSTM networks with attention mechanisms, have reported improved predictive performance compared to conventional recurrent models [14].

Building on these developments, the present study employs two complementary deep learning models: an ICEEMDAN–VMD–SMA–LSTM model optimized using a metaheuristic algorithm, and a CNN–LSTM model designed to capture spatial–temporal relationships between meteorological variables and groundwater dynamics. To further control model complexity and prevent overfitting, regularization techniques were incorporated within the deep learning architectures. Dropout layers with a rate of 0.2–0.3 were applied to reduce the co-adaptation of neurons, and early stopping was implemented based on validation loss to prevent excessive training. These mechanisms ensure that the models generalize well to unseen data despite the limited sample size. To enhance robustness and generalization capability, an Adaptive Weighting Model ensembles the outputs of individual models, dynamically adjusting their contributions based on predictive performance.

The Figure 5 depicts the model development and prediction workflow of the proposed hybrid framework. Feature sets derived from decomposed groundwater signals and meteorological variables are used to train a Slime Mould Algorithm–optimized LSTM (SMA–LSTM) model and a CNN–LSTM model. The individual model outputs are combined using an adaptive weighting model based on learnable coefficients, followed by fine-tuning to generate final groundwater level and quality predictions. All models were implemented using Python 3.12 (Python Software Foundation, Wilmington, DE, USA) with TensorFlow 2.16 and Keras 3.3 (Google LLC, Mountain View, CA, USA). Data preprocessing and analysis were performed using NumPy 1.26 and Pandas 2.2 libraries.

2.6.1. Long Short-Term Memory (LSTM) Model

Long Short-Term Memory (LSTM) networks are a class of recurrent neural networks specifically designed to learn long-term dependencies in sequential data. Groundwater level and quality time series often exhibit delayed responses to recharge, extraction, and climatic drivers, making LSTM architectures particularly suitable for this application.

The LSTM model employed in this study consists of memory cells with input, forget, and output gates that regulate information flow across time steps. This gated structure enables the model to retain relevant historical information while discarding noise and irrelevant fluctuations. The LSTM model serves as a baseline temporal predictor and provides a benchmark for evaluating the effectiveness of more advanced hybrid architectures.

2.6.2. ICEEMDAN–VMD–SMA–LSTM Hybrid Model

To improve forecasting performance under non-stationary conditions, a hybrid model combining signal decomposition, deep learning, and metaheuristic optimization was developed. Following ICEEMDAN and VMD, each intrinsic component was modeled independently using LSTM networks. This approach allows the model to focus on simplified temporal patterns at different frequency scales.

Hyperparameters of the LSTM networks, including learning rate, number of hidden units, and dropout ratio, were optimized using the Slime Mould Algorithm (SMA). The Slime Mould Algorithm (SMA) was configured with a population size of 30 and a maximum of 50 iterations. The optimization process aimed to minimize the validation error by tuning key LSTM hyperparameters, including learning rate, number of hidden units, and dropout rate. Convergence was determined based on the stabilization of fitness values or when the improvement between successive iterations fell below a threshold of 10⁻⁴. These settings were selected to ensure a balance between optimization efficiency and computational cost. SMA is a population-based optimization technique inspired by the foraging behavior of slime moulds and is well-suited for navigating complex, nonlinear search spaces. The optimized configuration enhances the convergence speed and predictive accuracy while reducing the risk of overfitting.

Metaheuristic optimization algorithms have been widely adopted in groundwater and hydrological modeling to address complex, high-dimensional search spaces where gradient-based methods are ineffective. Recent comprehensive reviews highlight the effectiveness of population-based metaheuristics, such as genetic algorithms, particle swarm optimization, and slime mould-inspired approaches, in balancing global exploration and local exploitation [15]. In this study, the Slime Mould Algorithm (SMA) was selected due to its adaptive feedback mechanism, dynamic parameter updating, and proven capability to efficiently tune deep learning hyperparameters under nonlinear and non-stationary conditions. This choice aligns with emerging trends in AI-driven hydrological modeling that emphasize robustness and convergence stability.

The

Table 2. Algorithmic workflow.

Step	Description
Input	Groundwater time series $X\left(t\right),$ meteorological data $M\left(t\right)$
Output	Predicted groundwater level $Y\left(t\right)$
1	Apply ICEEMDAN to X(t) to obtain intrinsic mode functions (IMFs)
2	For each IMF, apply Variational Mode Decomposition (VMD)
3	Obtain refined sub-signals VMDi
4	Construct feature set using lagged, rolling statistical, and time-based features
5	Initialize LSTM hyperparameters (learning rate, hidden units, dropout rate)
6	Optimize LSTM hyperparameters using Slime Mould Algorithm (SMA)
7	Train SMA-optimized LSTM model on decomposed components
8	Train CNN-LSTM model on meteorological feature representations
9	Generate predictions: YSMA from SMA-LSTM and YCNN from CNN-LSTM
10	Combine predictions using Adaptive Weighting Model: $Y\left(t\right)=\alpha \mathrm{Y}\_\mathrm{S}\mathrm{M}\mathrm{A}+ \beta \mathrm{Y}\_\mathrm{C}\mathrm{N}\mathrm{N}$
11	Fine-tune hybrid model parameters
12	Evaluate model using MAE, RMSE, R2, and NSE
13	Return predicted groundwater level $Y\left(t\right)$

describes the algorithmic workflow of the proposed ICEEMDAN–VMD–SMA–LSTM framework for groundwater level prediction. The algorithm applies ICEEMDAN and Variational Mode Decomposition to decompose groundwater time-series data, constructs engineered feature sets, optimizes LSTM hyperparameters using the Slime Mould Algorithm, and combines SMA–LSTM and CNN–LSTM predictions through an adaptive weighting model to generate final groundwater predictions evaluated using standard performance metrics.

2.6.3. Convolutional Neural Network–LSTM (CNN–LSTM) Model

To incorporate spatial variability and climatic influence, a hybrid CNN–LSTM model was developed. Convolutional neural networks are effective in extracting spatial features from gridded meteorological data, such as rainfall and temperature distributions. In the proposed framework, CNN layers first process meteorological inputs to learn the spatial patterns associated with recharge potential.

The extracted spatial features are then fed into an LSTM network to model temporal dependencies and long-term groundwater responses. This architecture enables the simultaneous learning of spatial and temporal relationships, improving the model’s ability to capture regional rainfall–groundwater interactions.

2.6.4. Adaptive Weighting Model (AWM)

Given that no single model consistently outperforms others under all hydrological conditions, an Adaptive Weighting Model (AWM) was implemented to ensemble predictions from multiple base models, including LSTM, CNN–LSTM, and ICEEMDAN–VMD–SMA–LSTM. The AWM assigns dynamic weights to individual model outputs based on their recent predictive performance, subject to a convexity constraint.

Weights are updated iteratively using error-based feedback, ensuring that models with lower prediction error contribute more significantly to the final forecast. This ensemble strategy enhances robustness, reduces model bias, and improves generalization across varying climatic and hydrological regimes.

2.6.5. Model Integration and Forecast Reconstruction

The final groundwater forecast is obtained by integrating outputs from all component models through the Adaptive Weighting Model. For decomposed signals, predictions from individual intrinsic components are aggregated to reconstruct the original groundwater time series. This hierarchical integration strategy ensures that both short-term variability and long-term trends are accurately represented in the final forecast.

The hybrid model development approach adopted in this study enables the effective handling of non-stationary groundwater dynamics and provides a flexible framework for incorporating additional data sources or modeling techniques in future extensions.

2.7. Model Training and Evaluation

Model training and evaluation were conducted to ensure that the proposed forecasting framework achieves high predictive accuracy, robustness, and generalization capability under varying hydrological and climatic conditions. It is important to note that projections under future climate scenarios are treated as scenario-based simulations rather than deterministic forecasts. While the proposed framework captures complex temporal relationships, it implicitly assumes that the learned relationships between climatic variables and groundwater responses remain structurally consistent over time. This assumption may be affected under strong non-stationary conditions induced by climate change. A walk-forward validation approach was adopted with a rolling window size of 60 months (5 years). At each step, the model was trained on historical data within the window and evaluated on the subsequent one-month prediction horizon. The window was then advanced sequentially across the dataset. This approach preserves temporal order and ensures robust evaluation under non-stationary conditions, closely simulating real-world forecasting scenarios. The walk-forward validation framework inherently mitigates overfitting by ensuring that model evaluation is performed on temporally unseen data. Unlike random train–test splits, this approach preserves chronological order and provides a more realistic assessment of predictive performance. The consistency of results across multiple rolling windows indicates that the model maintains stable performance and does not overfit to specific temporal segments. A structured training strategy, combined with rigorous evaluation metrics, was adopted to objectively assess model performance and reliability.

2.7.1. Training-Testing Strategy

The complete dataset was divided into training and testing subsets following a chronological split to preserve the temporal structure of groundwater time series. Historical data from 1994 to 2014 were used for model training, while data from 2015 to 2024 were reserved for testing and validation. This approach prevents information leakage from future observations and ensures a realistic evaluation of forecasting performance.

For deep learning models, a walk-forward (rolling-origin) validation strategy was employed during training to assess the sequential prediction capability. This method allows the models to be evaluated incrementally as new observations become available, closely mimicking real-world operational forecasting scenarios.

2.7.2. Model Training Configuration

All deep learning models were trained using consistent hyperparameter settings to enable fair comparison. Training was performed using adaptive optimization algorithms, including Adam and Slime Mould Algorithm-optimized learning rates for hybrid models. Typical training parameters included batch sizes of 32, training epochs ranging from 100 to 150, and dropout regularization to mitigate overfitting.

Activation functions were selected based on model architecture, with rectified linear units used in hidden layers and linear activation in output layers. Early stopping criteria were implemented based on validation loss to prevent excessive training and improve model generalization.

2.7.3. Evaluation Metrics

Model performance was evaluated using multiple statistical metrics that collectively assess accuracy, stability, and predictive reliability. These metrics include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of Determination (R²), and the Nash–Sutcliffe Efficiency (NSE). MAE and RMSE quantify the prediction error magnitude, while R² measures variance explanation and NSE evaluates the hydrological model efficiency.

2.7.4. Comparative Performance Assessment

To assess the effectiveness of the proposed hybrid framework, the performance of individual models—LSTM, CNN–LSTM, and ICEEMDAN–VMD–SMA–LSTM—was compared against the ensemble Adaptive Weighting Model. All models were evaluated on identical testing datasets using uniform metrics to ensure comparability.

The ensemble model consistently outperformed standalone architectures, demonstrating lower prediction errors and higher efficiency scores. This improvement highlights the benefits of combining signal decomposition, optimized deep learning, and adaptive ensemble learning for groundwater forecasting.

2.7.5. Robustness and Generalization Analysis

Model robustness was evaluated by examining performance across different hydrological conditions, including periods of high recharge and prolonged depletion. Sensitivity to extreme climatic events was assessed to ensure stability under non-stationary conditions. The results indicate that the ensemble framework maintains consistent performance and reduces error variance compared to individual models.

Overall, the adopted training and evaluation strategy ensures that the proposed framework provides reliable and generalizable groundwater forecasts suitable for decision-support applications under climate variability.

3. Results

3.1. Model Performance Evaluation

The predictive performance of four models—LSTM, CNN–LSTM, ICEEMDAN–VMD–SMA–LSTM, and the Adaptive Weighting Model (AWM)—was assessed using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of Determination (R²), and the Nash–Sutcliffe Efficiency (NSE). The baseline LSTM model captured general seasonal trends but exhibited higher error values under non-stationary conditions. The incorporation of spatial features through CNN–LSTM resulted in improved accuracy, while signal decomposition and metaheuristic optimization further enhanced the predictive performance.

The Adaptive Weighting Model achieved the best overall performance across all evaluation metrics, indicating superior accuracy and stability when combining multiple model outputs.

Table 3. Model performance evaluation.

Model	MAE (m)	RMSE (m)	${\mathit{R}}^2$	NSE
LSTM	0.412	0.671	0.881	0.864
CNN–LSTM	0.356	0.598	0.907	0.892
ICEEMDAN–VMD–SMA–LSTM	0.301	0.512	0.926	0.914
AWM (Optimized)	0.267	0.429	0.948	0.938

presents the quantitative comparison of groundwater forecasting models. A progressive improvement in prediction accuracy was observed from the baseline LSTM to the hybrid and ensemble-based models. The Adaptive Weighting Model achieved the lowest error values (MAE = 0.267 m, RMSE = 0.429 m) and the highest goodness-of-fit metrics (R² = 0.948, NSE = 0.938), demonstrating the effectiveness of integrating signal decomposition, optimized deep learning, and adaptive ensemble learning.

The grouped bar chart of Figure 6 presents the performance of LSTM, CNN–LSTM, ICEEMDAN–VMD–SMA–LSTM, and the proposed Adaptive Weighting Model (AWM) based on mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R²), and the Nash–Sutcliffe efficiency (NSE). The results demonstrate consistent performance improvement across successive model enhancements, with the AWM achieving the best overall accuracy.

3.2. Error and Goodness-of-Fit Analysis

This subsection jointly evaluates the prediction accuracy and goodness-of-fit to provide a consolidated assessment of model performance across error-based and efficiency-based metrics. Error-based evaluation highlights the advantages of hybrid and ensemble approaches. The AWM recorded the lowest MAE (0.267 m) and RMSE (0.429 m), reflecting reduced average deviation and improved handling of extreme fluctuations. Compared to the baseline LSTM model, the ensemble framework achieved an approximately 35% reduction in MAE along with a significant improvement in RMSE, demonstrating enhanced robustness under climatic variability and groundwater non-stationarity.

The ICEEMDAN–VMD–SMA–LSTM model also showed substantial error reduction relative to the standalone models, confirming the effectiveness of signal decomposition and hyperparameter optimization in isolating meaningful temporal patterns. Model reliability was further evaluated using R² and NSE metrics. All hybrid models achieved R² values greater than 0.90, indicating strong agreement between the predicted and observed groundwater levels. The Adaptive Weighting Model achieved the highest goodness-of-fit with R² = 0.948, explaining approximately 94.8% of the observed variance in groundwater levels.

Similarly, the NSE values exceeded the commonly accepted threshold of 0.75 for satisfactory hydrological model performance. The ensemble model achieved an NSE of 0.938, confirming excellent predictive efficiency and minimal deviation from the observed data.

3.3. Temporal Prediction Performance and Variable Relationships

Graphical comparisons between the observed and predicted groundwater levels were generated to visually assess model performance. Time-series plots indicate that the proposed models accurately replicated seasonal recharge cycles, long-term depletion trends, and short-term fluctuations. The ensemble model closely followed the observed groundwater behavior during both high-recharge and low-recharge periods, with minimal lag or amplitude distortion.

The visual agreement between the observed and predicted groundwater levels is further supported by the correlation structure among the groundwater and climatic variables. Correlation analysis revealed strong associations between groundwater levels and rainfall, as well as moderate correlations with temperature and salinity-related quality indicators. These relationships highlight the combined influence of climatic drivers and anthropogenic stressors on groundwater dynamics.

The Figure 7 depicts the Time-series comparison of the observed and predicted groundwater levels using the Adaptive Weighting Model during the testing period.

The Figure 8 depicts the Correlation analysis showing relationships between groundwater levels, meteorological variables, and groundwater quality parameters. The correlation analysis highlights significant relationships between the groundwater quality parameters and groundwater levels, indicating the influence of hydrochemical processes on groundwater dynamics. These insights support the selection of relevant features and enhance the interpretability of the modeling framework, although the primary prediction target remains the groundwater level.

3.4. Scenario-Based Groundwater Forecasting Under Climate Projections

The trained ensemble model was further applied to assess groundwater behavior under future climate scenarios derived from the CMIP6 projections. Scenario-based forecasts under the moderate (SSP2-4.5) and high-emission (SSP5-8.5) pathways indicate divergent groundwater trajectories. Under higher emission scenarios, the projected groundwater levels exhibit accelerated decline and increased variability, reflecting heightened climatic stress.

The Figure 9 is the projected groundwater level trends under the moderate-emission climate scenario using CMIP6 climate projections.

The Figure 10 is the projected groundwater level trends under the high-emission climate scenario, highlighting increased variability and long-term decline.

3.5. Summary of Results

Overall, the results demonstrate that the proposed hybrid deep learning framework significantly outperforms standalone models in forecasting groundwater levels under complex, non-stationary conditions. The Adaptive Weighting Model consistently achieved the highest accuracy, stability, and efficiency across all evaluation criteria, validating the effectiveness of integrating signal decomposition, optimized deep learning, and ensemble learning for groundwater forecasting applications.

Additional decomposition results for the groundwater level and water quality parameters are provided in the Appendix A (Figure A1, Figure A2, Figure A3, Figure A4, Figure A5 and Figure A6) for completeness.

4. Discussion

The results of this study were interpreted in the context of groundwater system dynamics, model design, and climate variability to provide deeper insights into the observed performance improvements. The proposed hybrid framework demonstrated a clear enhancement in predictive accuracy compared to standalone models, highlighting the importance of integrating advanced signal processing and deep learning techniques for modeling complex groundwater systems. The baseline LSTM model provided reasonable predictive performance, consistent with recent comparative analyses that report moderate accuracy for conventional recurrent neural networks in groundwater level prediction tasks [5]. The incorporation of convolutional layers in the CNN–LSTM model improved performance by capturing spatial patterns in meteorological inputs, highlighting the importance of spatial–temporal learning in climate-sensitive groundwater systems.

Further improvements achieved by the ICEEMDAN–VMD–SMA–LSTM model confirm the effectiveness of decomposition-based preprocessing in handling non-stationary groundwater time series. These findings are in agreement with recent studies that report enhanced performance when modal decomposition is combined with deep learning and ensemble strategies [1,7]. The proposed Adaptive Weighting Model achieved the highest predictive accuracy and efficiency, demonstrating that ensemble learning can effectively reduce model bias and improve robustness under varying hydrological conditions.

Recent advances in Transformer-based architectures have shown strong potential for long-horizon groundwater forecasting, particularly for capturing complex temporal dependencies over extended periods [10]. While such approaches were not explored in the present study, they represent a promising direction for future research.

From a management perspective, accurate and robust groundwater forecasts are essential for climate change adaptation and sustainable water resource planning. Integrating predictive models with broader surface–groundwater management frameworks has been identified as a key strategy for enhancing climate resilience and supporting informed decision-making [16]. The proposed framework provides a scalable and data-driven tool that can contribute to such integrated management efforts in groundwater-dependent regions.

Despite the relatively small dataset size, the combination of feature selection, regularization, and walk-forward validation ensures that the proposed model avoids overfitting and maintains generalization capability. The consistent performance across evaluation windows further supports the robustness of the framework.

4.1. Interpretation of Results

The improved performance of the proposed framework can be attributed to the synergistic integration of multi-stage signal decomposition, metaheuristic optimization, and ensemble learning. The ICEEMDAN–VMD effectively separates complex groundwater signals into distinct temporal components, enabling deep learning models to capture both short-term fluctuations and long-term trends with greater precision. This is particularly important in semi-arid regions where groundwater dynamics are influenced by irregular recharge patterns and anthropogenic extraction.

The incorporation of CNN–LSTM architecture enhances the model’s ability to capture spatial–temporal relationships between climatic variables and groundwater responses. Compared to the baseline LSTM model, the hybrid architectures demonstrated improved robustness under non-stationary conditions, confirming the importance of incorporating both spatial and temporal features in groundwater forecasting.

The application of the Slime Mould Algorithm for hyperparameter optimization further contributes to improved model performance by enhancing convergence efficiency and reducing the risk of overfitting. This is particularly beneficial in groundwater modeling, where datasets are often nonlinear, noisy, and influenced by multiple interacting variables.

The superior performance of the Adaptive Weighting Model indicates that combining complementary model strengths reduces individual model bias and improves generalization. The ensemble approach dynamically adjusts model contributions based on performance, allowing it to adapt to varying hydrological conditions. The observed reduction in prediction errors and improvement in R² and NSE values demonstrate that the proposed framework provides more stable and reliable forecasts compared to standalone and partially hybrid models.

These findings are consistent with recent studies that emphasize the advantages of decomposition-based and ensemble learning approaches in hydrological modeling. However, the present study extends existing work by integrating multiple advanced techniques into a unified framework specifically designed to handle climate-driven non-stationarity in groundwater systems.

While the proposed framework demonstrated superior performance compared to the LSTM and CNN–LSTM models, it is acknowledged that simpler machine learning approaches such as gradient boosting, support vector regression, and persistence-based forecasting can provide competitive baselines in certain scenarios. However, these models often struggle to capture complex non-stationary patterns and the multi-scale temporal dependencies inherent in groundwater systems.

The improved performance of the proposed framework justifies the additional computational complexity, particularly in applications where prediction accuracy and robustness under climate variability are critical. Nevertheless, future work may include comprehensive benchmarking against simpler models to further evaluate the trade-off between model complexity and performance.

4.2. Implications

From a practical perspective, the improved predictive accuracy of the proposed framework has significant implications for groundwater management in semi-arid and industrial regions. Reliable forecasting enables better planning of groundwater extraction, supports early warning systems for depletion, and facilitates climate-resilient water resource management strategies.

The integration of climate projections further enhances the applicability of the model for long-term decision-making. Scenario-based analysis indicates that groundwater levels are likely to decline under high-emission climate pathways, accompanied by increased variability. These findings highlight the urgent need for sustainable groundwater management practices, including controlled abstraction, artificial recharge, and improved industrial water reuse. The proposed framework can serve as a decision-support tool for policymakers and resource managers by providing accurate and timely groundwater forecasts under changing climatic conditions. Its adaptability also makes it suitable for application in other regions with similar hydrogeological characteristics, subject to appropriate data availability and model calibration.

Although the proposed model demonstrates strong predictive capability, forecasting under future climate scenarios involves inherent uncertainties. The use of CMIP6 projections introduces variability associated with climate model structure, emission scenarios, and downscaling techniques. In this study, these projections were used as representative scenarios rather than precise predictions. The assumption of temporal consistency between climatic drivers and groundwater response remains a limitation, particularly under rapidly changing hydroclimatic conditions. While decomposition-based modeling improves adaptability, it does not fully account for structural shifts in system behavior. Future research may focus on integrating multi-model climate ensembles and probabilistic forecasting approaches to explicitly quantify uncertainty and improve the robustness of long-term groundwater predictions.

While groundwater quality parameters were not directly modeled in the forecasting framework, their analysis provides valuable insights into hydrochemical interactions and system behavior. The limited temporal availability of quality data restricts their integration into long-term predictive modeling. Future research may extend the proposed framework to jointly model groundwater quantity and quality, enabling a more comprehensive assessment of groundwater system dynamics.

Although the proposed framework demonstrates strong predictive performance at the regional scale, spatial heterogeneity in hydrogeological conditions and groundwater abstraction patterns may influence localized model accuracy. The aggregation of well data may mask site-specific variations, particularly in hard-rock aquifer systems where groundwater behavior is highly heterogeneous. Future research may focus on well-specific modeling and spatial cross-validation to evaluate geographic variability in predictive performance. Incorporating spatially explicit modeling approaches or geostatistical methods could further enhance the robustness and applicability of groundwater forecasting frameworks in heterogeneous regions.

The findings of this study are consistent with prior research highlighting the limitations of purely regulatory approaches in groundwater management. As noted by Liptrot and Hussein [11], policies such as groundwater pricing and well licensing frequently encounter resistance from agricultural stakeholders, limiting their effectiveness.

In this context, the proposed predictive framework can serve as a complementary decision-support tool, enabling more informed and adaptive groundwater management strategies. This suggests that integrating data-driven forecasting models with governance mechanisms can enhance the effectiveness of groundwater regulation. Particularly in regions with competing water demands, such integrated approaches can support sustainable allocation and reduce the risk of overexploitation.

4.3. Limitations and Future Research Directions

While the proposed framework demonstrates strong predictive performance, certain limitations remain. Forecast accuracy depends on the quality and spatial resolution of the input datasets, particularly climate projections. In addition, model transferability to regions with distinct hydrogeological characteristics may require retraining with localized data. Future research may explore the integration of Transformer-based architectures, uncertainty quantification methods, and coupled surface–groundwater models to further improve long-horizon forecasting under climate change.

Although CMIP6 climate projections provide valuable insights into potential future hydroclimatic conditions, uncertainties remain due to model structure, emission scenario assumptions, and downscaling processes. In particular, differences among global climate models and the spread between emission pathways introduce variability in projected rainfall and temperature inputs, which may propagate into groundwater forecasts. Consequently, the scenario-based results presented in this study should be interpreted as indicative trends rather than deterministic predictions.

The improved predictive performance obtained in this study is consistent with previous research demonstrating the benefits of metaheuristic-assisted groundwater modeling. Similar improvements in prediction accuracy have been reported when optimization algorithms were integrated with groundwater prediction frameworks to enhance parameter calibration and model convergence [9]. The results of the proposed hybrid framework demonstrate improved predictive performance compared with conventional machine learning models. Similar findings have been reported in previous studies where machine learning models combined with geostatistical techniques achieved reliable groundwater level predictions across heterogeneous aquifer systems [6]. Beyond hydroclimatic drivers, groundwater scarcity is also shaped by socio-economic and governance factors. Previous studies have emphasized that water scarcity is often linked to patterns of resource governance, industrial demand, and policy decisions rather than purely natural limitations [17]. In regions with intensive industrial activity, such as Tiruppur, these socio-economic pressures can significantly influence the groundwater depletion trends. Water scarcity is increasingly recognized as a complex socio-environmental phenomenon shaped by both natural and institutional factors. Studies examining water governance and policy narratives have shown that perceptions of scarcity are often influenced by institutional framing, resource management practices, and policy priorities rather than solely hydrological constraints [18]. Such perspectives highlight the importance of integrating scientific forecasting with sustainable water management strategies. Future studies may incorporate nitrate and other trace contaminants where consistent long-term datasets are available.

5. Conclusions

This study developed a hybrid artificial intelligence-based framework for groundwater level and quality forecasting under climate-driven non-stationary conditions by integrating multi-stage signal decomposition, optimized deep learning models, and adaptive ensemble learning. The proposed ICEEMDAN–VMD–SMA–LSTM–CNN–LSTM framework effectively captures complex temporal and spatial dependencies in groundwater systems influenced by climatic variability and anthropogenic pressures.

The results demonstrate that the proposed framework significantly outperforms standalone and partially hybrid models. The Adaptive Weighting Model achieved the best predictive performance, with a mean absolute error (MAE) of 0.267 m, root mean square error (RMSE) of 0.429 m, coefficient of determination (R²) of 0.948, and Nash–Sutcliffe efficiency (NSE) of 0.938. These results confirm the effectiveness of integrating signal decomposition, metaheuristic optimization, and ensemble learning for improving forecasting accuracy and reducing residual variability.

The study also highlights that multi-stage decomposition plays a critical role in handling non-stationary groundwater signals, while ensemble learning enhances robustness by combining complementary model strengths. The incorporation of climate projections further demonstrates the capability of the proposed framework to assess future groundwater behavior under different emission scenarios, indicating increased groundwater stress and variability under high-emission conditions.

From an application perspective, the proposed framework provides a reliable and scalable tool for groundwater resource management in semi-arid and industrial regions. It supports informed decision-making by enabling accurate forecasting, early warning of groundwater depletion, and climate-resilient planning strategies.

Overall, the findings of this study contribute to advancing data-driven groundwater forecasting methodologies by providing an integrated, robust, and adaptable framework capable of addressing the challenges posed by climate variability and complex hydrogeological conditions.

However, the projections presented in this study should be interpreted as scenario-based assessments rather than precise forecasts, as uncertainties associated with climate models and non-stationary system behavior are not explicitly quantified. Future work may incorporate uncertainty propagation techniques and ensemble climate modeling to enhance prediction reliability.

Future studies may incorporate additional baseline models, including gradient boosting and support vector regression, to further assess model efficiency and comparative performance, and may focus on extending the framework to integrate groundwater quality forecasting alongside groundwater level prediction, subject to the availability of long-term datasets. It may also include spatially explicit modeling and well-level validation to capture localized groundwater dynamics and improve model generalization across heterogeneous aquifer systems.

Furthermore, sustainable groundwater management requires the integration of predictive modeling with effective regulatory frameworks, as emphasized in recent policy-oriented studies [11].