Artificial Intelligence Advancements for Accurate Groundwater Level Modelling: An Updated Synthesis and Review

Abstract

Artificial Intelligence (AI) methods, including Artificial Neural Networks (ANNs), Adaptive Neuro-Fuzzy Inference Systems (ANFISs), Support Vector Machines (SVMs), Deep Learning (DL), Genetic Programming (GP) and Hybrid Algorithms, have proven to be important tools for accurate groundwater level (GWL) modelling. Through an analysis of the results obtained in numerous articles published in high-impact journals during 2001–2023, this comprehensive review examines each method’s capabilities, their combinations, and critical considerations about selecting appropriate input parameters, using optimisation algorithms, and considering the natural physical conditions of the territories under investigation to improve the models’ accuracy. For example, ANN takes advantage of its ability to recognise complex patterns and non-linear relationships between input and output variables. In addition, ANFIS shows potential in processing diverse environmental data and offers higher accuracy than alternative methods such as ANN, SVM, and GP. SVM excels at efficiently modelling complex relationships and heterogeneous data. Meanwhile, DL methods, such as Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNNs), are crucial in improving prediction accuracy at different temporal and spatial scales. GP methods have also shown promise in modelling complex and nonlinear relationships in groundwater data, providing more accurate and reliable predictions when combined with optimisation techniques and uncertainty analysis. Therefore, integrating these methods and optimisation techniques (Hybrid Algorithms), tailored to specific hydrological and hydrogeological conditions, can significantly increase the predictive capability of GWL models and improve the planning and management of water resources. These findings emphasise the importance of thoroughly understanding (a priori) the functionalities and capabilities of each potentially beneficial AI-based methodology, along with the knowledge of the physical characteristics of the territory under investigation, to optimise GWL predictive models.

1. Introduction

Groundwater, a crucial global hydric resource, profoundly impacts various aspects of human life, including agriculture, industrial development, and drinking water provision [1,2,3]. The ability to comprehend and forecast the fluctuations in groundwater level (GWL) is paramount for effective water resource planning and management [4]. However, conventional statistical methods in GWL predictive modelling have limitations, such as data dependency and complex calibration processes, requiring the use of advanced and more robust approaches like the application of Artificial Intelligence (AI) methods [5,6].

Advancements in technology and modern methods have enabled AI to significantly enhance the prediction, planning, and management of groundwater resources [7]. Scientific investigation in this field can help identify the best strategies for the conservation and optimal management of this resource in the face of current and future changes, including climate change [8]. Furthermore, these studies can contribute to developing new technologies for more accurate monitoring and measurement of groundwater tables, providing essential information for better decision-making policies [9]. Hence, groundwater research is of great scientific and technical value and is crucial in conserving groundwater resources and formulating appropriate management strategies [10].

Over the past twenty-two years, AI has been used extensively to overcome the limitations of conventional statistical approaches in simulating GWL. Indeed, based on the compilation and analysis of several articles published between 2001 and 2023 (see references list), it is evident that different research groups are increasing interest in using various AI-based methods to produce GWL predictive models. However, it is also evident that recent studies are lacking in many specific countries (Figure 1). This could be due to the absence of demand for groundwater resources and limited knowledge about the potential and procedures required for implementing AI-based predictive modelling that may influence this trend.

Figure 2 shows a growing tendency in recent years to use newer AI-based methods, such as Artificial Neural Networks (ANNs), Deep Learning (DL), Genetic Programming (GP), and their combination with Hybrid Algorithms for more accurate GWL studies.

This work comprehensively reviews the AI-based methods most used in recent years to predict GWL. Its primary purpose is to highlight each method’s capabilities, combinations, and critical considerations for improving accuracy/efficiency in GWL prediction models. It also presents potential solutions to the challenges associated with their use. Practically, this review can assist water science researchers, such as geographers, geologists, engineers, environmentalists, and other professionals, to make methodological choices when modelling GWL and managing the related natural resources. Additionally, this article evaluates Hybrid Algorithms that combine AI and statistical methods, examining their applications and advantages in geographic territories with limited environmental data. Ultimately, this review promotes awareness of advanced AI-based methods in groundwater management, emphasising practical applications and future potential.

2. Methodology

This study’s methodology involved synthesising and analysing data from more than a hundred articles published between 2001 and 2023. The goal was to identify trends in AI utilisation across different research groups and countries. Critical factors include the types of AI-based methods used (ANN, ANFIS, SVM, DL, GP, and Hybrid Algorithms), the simulated parameters (e.g., water table fluctuations or aquifer recharge rates), and the geographical contexts of the realised investigations. This analysis will provide insights into how effective AI is in improving groundwater assessments, identifying potential barriers to AI adoption (such as technological readiness or institutional constraints), and offer future research directions.

3. Artificial Intelligence Methods for GWL Modelling

3.1. Artificial Neural Networks

ANNs are some of the most powerful machine learning algorithms based on the structure and function of the human brain’s neural system. These networks consist of neurons and their connections [11]. Each neuron in this network acts as a processing unit, with the weights corresponding to its connections determining the importance and impact of the inputs on the outputs [10]. The structure of an ANN typically includes three layers [12]: (1) the input layer is responsible for receiving the input variables; (2) the hidden layers, situated between the input and output layers, are responsible for the main information processing; and (3) finally, the output layer produces the result. The working method of ANNs in the training process involves providing inputs to the network, which produces corresponding outputs. The generated error is then compared with the desired output, and the network’s weights are automatically adjusted to minimise the error. This process is repeated until the weights reach an optimal value, allowing the network to predict more accurately [13].

Figure 3 illustrates the classical structure of an ANN. It also shows the connections between the neurons, including all connections between the neurons of the input layer and the hidden layer and between the neurons of the hidden layer and the output layer. These connections represent the weights optimised during network training to provide the best result.

In groundwater studies, ANNs are powerful tools with many applications [15]. These include GWL prediction, identifying geologic patterns associated with groundwater, and improving the performance of existing groundwater potential mapping. ANNs can improve cartographic models, increasing their predictive accuracy [16]. An analysis of 41 articles from 2001 to 2022 shows that ANNs are widely used worldwide, with India and Iran standing out as essential players (Figure 4).

Ref. [17] conducted a study in the Yamuna–Hindon Inter basin, India, and found that the ANN Levenberg–Marquardt (LM) algorithm was slightly superior to that of Bayesian Regularisation (BR) and Back Propagation (BP) algorithms in GWL prediction accuracy. The input data included the maximum depth of the water table for agricultural purposes. Similarly, ref. [18] in Jaipur and Rajasthan, India, demonstrated that an ANN outperformed time series forecasting methods (e.g., ARIMA) in describing GWL and water quality parameters. The input data for this study included water levels.

In [19], an investigation conducted in Tikri Kalan, West Delhi, India, a 3-15-1 ANN architecture was highly effective in predicting GWL fluctuations. Input data included rainfall, land use, and population. Likewise, [12] in Pakistan successfully used an ANN to predict daily GWL using input data such as rainfall, maximum and minimum temperatures, solar radiation, humidity, wind speed, and terrain elevation. A study by [8] in the Nabhana watershed in Tunisia showed that monthly GWL was influenced by rainfall, evaporation, and previous values. The input data included rainfall, previous groundwater levels, and evaporation. Ref. [11] Aquifer, South Khorasan, Iran, demonstrated that a Bayesian Network (BN) performed better than an ANN. The input data included monthly rainfall, temperature, aquifer recharge/discharge, and piezometric data. The multi-layer Feed-Forward Network was identified as the best method in a study by [20] in Montgomery County, PA, USA. The input data included humidity, rainfall, and temperature. Similarly, ref. [15] in Odisha, India, showed that evaporation and runoff were critical factors in predicting GWL. A Recurrent Neural Network (RNN) was used, with input data including rainfall, temperature, humidity, runoff, and evaporation. In South Korea, ref. [21] showed that a Feed-Forward Artificial Neural Network (FFANN) successfully predicted hourly GWL. The input data included surface water levels and groundwater extraction parameters. Ref. [22], in the Shamil–Ashtor Plain, Iran, demonstrated that an Extreme Learning Machine (ELM) was significantly more accurate than a Radial Basis Function Neural Network (RBFNN), Multiple Linear Regression (MLR), and Autoregressive Moving Average (ARMA) in predicting monthly GWL. The input data included hydrological and climatic data. Ref. [23], in Southwest Germany, conducted a study where the nonlinear autoregressive network (NARX) effectively predicted GWL in wells unaffected by human activities. The input data included rainfall and temperature. Ref. [21], in Longyan City, Fujian Province, China, showed that an RBFNN was more efficient in predicting monthly GWL. The input data included daily groundwater levels. In Canada, [24] showed that the ELM outperformed the Support Vector Regression (SVR) method in predicting monthly GWL. The input data included meteorological and hydrological data. Ref. [25], in Taiwan, demonstrated that a Feed-Forward Back Propagation Neural Network (FFBPNN) effectively simulated GWL fluctuations. The input data included daily groundwater levels. Ref. [26], in Jenderam Hilir, Selangor, Malaysia, showed that the multi-layer perceptron (MLP) successfully predicted GWL in tropical areas. The input data included hydrological and climatic parameters. Table S1 (Supplementary Materials) summarises the evaluated articles that used ANN methods and draws a general conclusion.

3.2. Adaptive Neuro-Fuzzy Inference Systems

An ANFIS is a powerful hybrid computational algorithm that integrates the principles of fuzzy logic and neural networks to effectively address complex system modelling and decision-making tasks [27,28]. It combines the qualitative reasoning of fuzzy logic with the learning capabilities of neural networks, enabling it to handle uncertain and nonlinear relationships in data [29].

An ANFIS typically consists of five layers (Figure 5), each serving a specific function in the inference process [28,30,31]:

• Layer 1 (input layer)—represents the system’s input variables. Each node in this layer applies a membership function to the input data, converting crisp inputs into fuzzy linguistic terms. The membership functions define the degree to which each input belongs to different fuzzy sets;
• Layer 2 (fuzzy layer)—nodes in this layer calculate the firing strength of each rule. Each node computes the degree of matching between the input variables and the fuzzy sets defined in the antecedents of the fuzzy rules. The firing strength represents the activation level of each rule based on the input data;
• Layer 3 (normalisation layer)—the firing strengths computed in Layer 2 are normalised to ensure that they sum up to one. Normalisation is necessary to maintain consistency in the aggregation process and prevent biases due to varying firing strengths;
• Layer 4 (consequent layer)—computes each rule’s output based on its firing strengths and consequent parameters. It combines the normalised firing strengths with consequent parameters to generate each rule’s local outputs;
• Layer 5 (output layer)—nodes in this layer aggregate the local outputs of Layer 4 to produce the overall output of the ANFIS. The final output is obtained through a weighted summation of the local outputs, representing the model’s prediction or decision.

The mathematical formulation of an ANFIS involves a set of equations corresponding to each layer, including equations for membership function computation, firing strength calculation, normalisation, and output aggregation [32,33]. These equations encapsulate the underlying fuzzy logic and neural network principles employed by the ANFIS [34].

ANFISs have emerged as versatile tools with extensive applications across various domains in groundwater sciences. One prominent application of ANFISs is in GWL prediction, where they utilise historical data encompassing factors such as precipitation, temperature, soil characteristics, and land use patterns [35]. By analysing these data inputs, ANFISs can accurately forecast future groundwater levels, aiding in effective water resource management and planning [36]. Furthermore, ANFISs are crucial in assessing groundwater quality [37]. By correlating water parameters, such as pH, dissolved oxygen, and contaminant concentrations, with environmental variables and human activities, ANFISs enable comprehensive evaluations of groundwater quality. This capability is invaluable for identifying potential contamination sources, assessing environmental impacts, and implementing remediation strategies [38]. Additionally, ANFISs identify hydrological patterns and trends [39]. By analysing complex datasets and capturing nonlinear relationships, ANFISs help uncover underlying patterns in groundwater dynamics, including seasonal fluctuations, recharge rates, and flow pathways. This information is instrumental in making informed decisions about water allocation, pollution control, and sustainable development [40]. ANFISs are powerful for groundwater modelling and prediction in diverse environmental settings.

This study analysed 18 articles published in reputable journals between 2009 and 2023 and one published in 2024. Based on these articles, most studies were performed in Iran and the United States (Figure 6).

Ref. [38] conducted studies in the Pravara River, India, using the ANFIS method to predict GWL. These studies demonstrate that ANFISs outperform ANNs, and the spatial distribution data of Inverse-Distance Weighting (IDW) are also reliable. Ref. [34] examined the accuracy of the Improved Simulated Annealing (ISA) algorithm and Least Squares Support Vector Machine (LSSVM) in predicting GWL in Iran. They indicated that the ISA-LSSVM has a higher accuracy. Research conducted by [41] comparing the accuracy of the Neuro-Fuzzy-based group method of data handling (NF-GMDH) improved through Particle Swarm Optimisation (PSO) algorithms with other methods in predicting GWL in the United States shows that NF-GMDH-PSO outperforms other methods and employs innovative modelling approaches. Evaluated studies by [42] in South Korea, assessing the acceptability of the ANFIS and Stochastic Coded Federated Learning (SCFL) methods, demonstrate that the ANFIS method is acceptable. Ref. [40] evaluated the accuracy of ANFIS-FCM (fuzzy c-means) using sym4 mother wavelets in predicting GWL in Iran. They showed that this achieved higher accuracy and benefits from their combination. Evaluated studies by [39], comparing the accuracy of the ANFIS with other methods in predicting GWL in China, demonstrate that ANFIS has higher accuracy than others, such as RBFNN and Grey Self-Memory (GSM). Ref. [43] compared the accuracy of Gaussian Process Regression (GPR) with ANFIS’s in predicting multi-step monthly GWL in Sullia Taluk, India. They found that GPR performed better than ANFIS. Ref. [35], by comparing the accuracy of the ANFIS with other methods in predicting monthly GWL fluctuations in Kashan Plain, Iran, demonstrate that the ANFIS method (based on the orthogonal mean shape) has higher accuracy. Evaluated studies by [33] in the Langat Basin, Malaysia, comparing the accuracy of ANFIS with other methods in predicting GWL, show that the ANFIS method has higher accuracy than others. Table S2 (Supplementary Materials) summarises the evaluated articles that used ANFIS methods and draws a general conclusion.

3.3. Support Vector Machines

SVMs, which also belong to the family of machine learning tools, have shown remarkable success in various hydrogeological applications, including GWL modelling [44,45,46]. SVMs create an optimal hyperplane that classifies data points into different classes or predicts GWL, maximising the distance between classes to improve robustness and generalisation [47,48]. SVMs are characterised by processing high-dimensional data and capturing complex patterns within the dataset, considering multiple influencing factors simultaneously and providing a holistic perspective on groundwater dynamics [49]. Their adaptability to various kernel functions, such as the RBF and polynomial kernels, further enhances their ability to model complex relationships within hydrogeological systems [50,51,52]. SVMs have been used to address prediction accuracy and uncertainty challenges, demonstrating robust performance compared to conventional methods and optimising water resource management and environmental planning [53,54]. With the advances in machine learning, SVMs remain a valuable tool for predicting GWL and improving our understanding of subsurface hydrological processes [48]. In summary, the classical structure of an SVM is as follows [49,50] (Figure 7):

• Learning phase—the SVM algorithm seeks to find the optimal hyperplane separator between different classes in the training dataset;
• Choosing the SVM type—select the type of SVM among linear and nonlinear models and specify the kernel function to transform data into a higher-dimensional space if necessary;
• Determining hyperparameters—these are essential parameters such as kernel function and objective function like C for controlling the error rate and the number of misclassifications;
• Training—involves finding the SVM model using the training dataset and determining the optimal parameters;
• Evaluation—involves applying the model to test data to assess its performance and measuring evaluation metrics such as accuracy and confusion matrix;
• Tuning—after evaluating the model, we may need to re-adjust parameters or change the type of kernel function to improve model performance;
• Utilisation—the trained model is used to predict the class label of new data;
• Maintenance and update—if the model needs updating (e.g., with new data), it may need to be retrained or updated.

This study analyses 14 articles published between 2011 and 2023, providing an assessment of each article followed by a general conclusion drawn from them (Figure 8).

In the studies conducted by [3], a combined Wavelet Transform (WT) LSSVM was employed to predict GWL in the Saveh region, Iran. This combination performed better than independent methods using precipitation, evaporation, temperature, and GWL as input data. For instance, on specific days with high precipitation and low temperature, the WT-LSSVM could predict GWL with higher accuracy. Ref. [7] investigated the performance of an SVM for predicting GWL in the Azerbaijan region, Iran. The results indicate that the SVM outperformed other ANNs. For example, during the summer season, with increasing temperatures, the SVM can predict GWL variations with higher accuracy. A paper published by [55] examined the performance of the SVM-ALO (ant lion optimiser) in estimating monthly water levels in the Purbamedinipur region, India. The results show that this method performed very well. For instance, with seasonal variations in precipitation and temperature, the SVM-ALO could predict GWL changes with higher accuracy. Ref. [45] studied the GTB method’s performance in predicting GWL one month ahead in the Karnataka region, India. The results indicate that the Gradient Tree Boosting (GTB) outperformed the ANFIS and Group Method of Data Handling (GMDH). For example, the GTB could predict GWL variations more accurately during rainy seasons with heavy rainfall. Ref. [44] investigated the performance of the Multi-Objective Genetic Algorithm and Support Vector Machine (MOGA-SVM) algorithm in predicting GWL fluctuations with different lead times in Taiwan. The results showed that this model outperformed the Self-Organisation Maps (SOM)-SVM algorithm. For instance, with seasonal and yearly variations in precipitation and evaporation, the MOGA-SVM can predict GWL changes with greater accuracy. Ref. [50] examined the performance of the Radial Basis Function–Support Vector Regression (RBF-SVR) in predicting GWL in various regions of the United States. The results indicate that this method performed better, especially in different seasons. For example, the RBF-SVR could predict GWL more accurately in areas with variable precipitation. In a study conducted by [56], the performance of two methods, SVR (Polynomial) and M5 Model Tree (M5Tree), in predicting monthly GWL changes in Ardabil Plain, Iran, was examined. The results showed that both models accurately predicted monthly changes. For example, the M5Tree could predict GWL more accurately during cold seasons with low-temperature fluctuations. Ref. [54] investigated the impact of variations in Total Water Storage (TWS) on GWL in India. The results indicate that TWS was identified as an essential variable, and the SVR method better predicted GWL than others. The research conducted by [46] examined the performance of RBF-SVR in predicting GWL in the Indo-Gangetic Basin, India. The results show that this method performed better for improved prediction and utilised satellite data as inputs. Ref. [48] investigated the performance of a Least-Squares Support Vector Machine (LS-SVM) in predicting hourly fluctuations in GWL in England. The results indicated that this method provided superior performance in predicting these fluctuations. Ref. [49] examined the performance of RBF-SVM in predicting monthly fluctuations in GWL in China. The results showed that this method is more accurate and less uncertain, using precipitation, evaporation, and temperature as input variables. For example, in regions with hot and dry climates, RBF-SVM can predict changes in GWL more accurately. Ref. [53] compared the performance of ANNs, SVMs, and ANFISs in predicting monthly fluctuations in GWL in Florida, United States. Table S3 (Supplementary Materials) summarises the evaluated articles that used SVM methods and draws a general conclusion.

3.4. Deep Learning

Deep Learning, including ensemble-based XGBoost (eXtreme Gradient Boosting) algorithms, self-attention-integrated Long Short-Term Memory (LSTM), and hybrid LSTM methods, are increasingly utilised for GWL prediction. Indeed, [5] employed ensemble-based XGBoost in the Sacramento River Basin, prioritising well-specific features for accuracy. Ref. [57] integrated self-attention mechanisms with LSTM in Iran, outperforming other methods with a self-attention (SA) temporal convolutional network (SATCN)-LSTM when capturing dependencies. [6] utilised LSTM in Northern France, proficiently filling data gaps and representing fluctuations. LSTM, in combination with the Lion Algorithm [58], excelled in GWL forecasting, considering regional complexities. [59] applied LSTM in Jeju Island, Korea, to predict GWL. Figure 9 shows an example of a DL flowchart for predicting the GWL.

This study analysed 12 articles published between 2018 and 2023, providing an assessment of each article followed by a general conclusion drawn from them (Figure 10).

Ref. [60] investigation shows that a robust cost function with the least trimmed squares and the Whittaker smoother effectively controls outliers and noise. This model uses GWL data affected by outliers and noise in South Korea. According to a study by [61], the LSTM accurately predicts daily GWL. This model used daily precipitation data on Jeju Island, Korea. Research by [62] recommended the DL for predicting groundwater depth, demonstrating superior accuracy compared to ELM and GPR. This model uses precipitation, river stage, temperature, recharge, and groundwater depth data in the Konan Basin, Kochi Prefecture, Japan. According to a study by [63], LSTM networks outperform RNN in predicting groundwater table responses to storm events. This model uses groundwater table, rainfall, and sea level data in Norfolk, Virginia, USA. Research by [64,65] shows that Convolutional Neural Network (CNN)-based Deep Learning methods enhance the prediction of terrestrial Total Water Storage anomalies and improve model performance. This model used GRACE Satellite Data in India. According to a study by [66], comprehensive data preprocessing enhances predictive model accuracy in detecting hydrologically stressed aquifer conditions. These include NARX-DNN, LSTM, Gated Recurrent Unit (GRU), Autoregressive and Exogenous (ARX) methods, and water table level time series data in Pohang Gibuk, Korea, which are used. Research by [67] shows that the proposed model outperformed a Feed-Forward Neural Network (FFNN) and Double-LSTM in predicting water table depth. This method used monthly water diversion, evaporation, precipitation, temperature, and time data from the Hetao Irrigation District, China. Table S4 (Supplementary Materials) summarises the evaluated articles that used DL methods and draws a general conclusion.

3.5. Genetic Programming

GP is an AI-based method that uses algorithms inspired by natural selection to solve problems [4,68]. GP is a subset of Genetic Algorithms and is often used to generate computer programmes, especially in domains where traditional algorithms may not perform well [2,69]. In GP, codes are generated in a structured manner [70]. These structures are called “species” or “individuals” that genetic algorithms improve. Typically, a programme in GP is represented as a structured tree consisting of operators, operands, and variables [71]. Each node in this tree can represent an operator or an operand [72]. GP is applied in groundwater studies to model hydrological processes, optimise pumping strategies, predict GWL, identify influential factors, and design monitoring networks [68]. The structure of GP typically involves a tree-based framework, where various parameters, operators, and functions are arranged within nodes connected by branches. Each node in the tree comprises two distinct sets, i.e., a function set and a terminal set [2]:

• Function set—nodes within the function set encapsulate mathematical functions like square, sine, and tangent; arithmetic operators such as addition, subtraction, and division; Boolean operators like AND and OR; and additional user-defined expressions; these nodes usually have child arguments.
• Terminal set—nodes belonging to the terminal set consist of numerical constants and variables, representing inputs or random variables within the problem domain.

Figure 11 shows an example of a GP model consisting of two tree-branch structures [2]. The left tree structure is (7 − X1) × (X2 + 5), where X1 and X2 symbolise random variables. The right tree structure can be constructed as sqrt(Y1) + (X1/8), where Y1 denotes another random variable. The GP’s tree-based configuration essentially entails nodes embodying functions and terminals, structured hierarchically to formulate mathematical expressions or programmes to address specific problem-solving objectives.

This study analyses 9 articles published between 2011 and 2023, providing an assessment of each article followed by a general conclusion drawn from them (Figure 12).

Based on the research conducted by [4], Gene Expression Programming (GEP), particularly its three-gene structure, proves highly effective for GWL modelling. By leveraging time series data, this method captured the temporal variations and complexities inherent in the groundwater data of Urmie Lake, Iran. This study underscores the potential of GEP for providing accurate and reliable GWL predictions, which is essential for effective water resource management in the region. According to the investigations by [2], GP offers precise GWL predictions, with a notable emphasis on the strategic placement of weather stations. The study reveals that accurate weather station locations significantly enhance the model’s predictive capability. Ref. [70] research highlights that Gaussian Process Regression (GPR) surpasses GP in predicting groundwater salinity. The comparison demonstrates that GPR offers better precision and reliability in handling the complex relationships between pumping activities and salinity concentrations. In 2017, [73] introduced novel formulations integrating uncertainty into simulation and optimisation processes, showcasing improved accuracy and reduced computational time. Their method utilises GP combined with Continuous Ant Colony Optimisation and Monte Carlo simulation to handle various hydrological parameters. According to the research by [71], incorporating uncertainty quantification into GP-based methods is essential for refining GWL predictions. This study uses observations from three wells to illustrate how uncertainty analysis can enhance the reliability and robustness of GWL models. A study by [68], an adaptive GP-based method for Bayesian experimental design, proved efficient in identifying groundwater contaminant sources. By integrating Markov Chain Monte Carlo (MCMC) methods, this study enhances the accuracy and efficiency of environmental monitoring efforts. Based on the findings of [72], GP outperforms the ANFIS in predicting monthly GWL, mainly when using precipitation and evaporation data. This research highlights GP’s capability to capture the nonlinear relationships between climatic variables and GWL. Research by [69] demonstrates that GEP outperforms other AI-based methods and the ARMA in short-term GWL forecasting. By utilising daily GWL and rainfall data, GEP provides a robust and adaptable approach for dynamic and nonlinear groundwater systems. According to a study by [74], linear regression models are proficient in predicting summer GWL, whereas ANN is more effective in making winter predictions. This seasonal differentiation suggests that the choice of prediction models should consider seasonal variations to optimise accuracy. This study used local weather, lake stage, and stream flow data, demonstrating the importance of contextual inputs in GWL modelling for the Searsville basin. Table S5 (Supplementary Materials) summarises the evaluated articles that used GP methods and draws a general conclusion.

3.6. Hybrid Algorithms

Integrating AI with hybrid optimisation algorithms has proven effective in overcoming challenges associated with traditional methods in GWL modelling [75]. Notable achievements include the development of robust methods enhancing accuracy, particularly in arid and semi-arid regions [1,76]. Incorporating hybrid evolutionary algorithms, including the Non-Dominated Sorting Genetic Algorithm (NSGA)-II, Multi-Objective Particle Swarm Optimisation (MOPSO), and Whale Algorithm, contributes to model resilience, addressing issues like over- and underestimation [77,78]. Advanced methods improve prediction accuracy and prove resource-efficient, especially in areas with limited statistical data or challenges in employing complex numerical models [79,80]. This study analyses 35 articles published between 2010 and 2023, providing an assessment of each article followed by a general conclusion drawn from them (Figure 13).

Ref. [76] Introduced a novel multi-layer perceptron learning process, which included a Non-Dominated Sorting Genetic Algorithm (2NSGA II) and Multi-Objective Particle Swarm Optimisation (MOPSO) for GWL prediction in Yazd, Iran. By merging multiple evolutionary algorithms, including NSGA II and MOPSO, the researchers aimed to enhance prediction accuracy. The input variables comprised temperature, precipitation, and GWL datasets. This research emphasises the significance of tailored modelling for addressing water resource challenges in arid regions. Insights from this study pave the way for advancing groundwater management practises in similar climatic conditions. Ref. [3] investigated hybrid wavelet–machine learning methods for GWL prediction in Saveh, Iran, focusing on the WT-LSSVM. Through extensive experimentation with various input parameters like precipitation, evapotranspiration, temperature, and groundwater level datasets, this study demonstrates the superior performance of the WT-LSSVM model inaccurate forecasting. Advanced modelling methods enhance water resource sustainability and resilience in similar geographic contexts. Ref. [79] present the outlier-robust Extreme Learning Machine (ORELM) method as a promising GWL prediction alternative. Through comparative analysis with conventional methods like the Groundwater Model System (GMS), Genetic Algorithm (GA)-ANN, Imperialist Competitive Algorithm (ICA)-ANN Hybrid Algorithms, ELM, and ORELM, the research highlights the ORELM model’s superior accuracy, efficiency, and cost-effectiveness. Ref. [1] contribute to groundwater research by establishing the effectiveness of Hybrid Algorithms for predicting GWL. Focusing on the Shazand Plain, Iran, this study compares various hybrid methods such as PSO-ANN, ACA-ANN, LSTM, LS-SVM, ORELM, and the GMS. By utilising monthly GWL datasets and advanced approaches, this research underlines the importance of adopting tailored methodologies for accurate groundwater forecasting. Ref. [80] introduce the Augmented Artificial Ecosystem Optimisation-based Multi-Layer Perceptron (AAEO-MLP) for monthly GWL forecasting. Situated in Karnataka, India, this study showcases the model’s superior accuracy, convergence, and stability performance compared to other forecasting methods. Ref. [81] investigate the ANFIS Grasshopper Optimisation Algorithm (GOA) for GWL prediction, focusing on the Ardebil Plain, Iran. By integrating optimisation algorithms with ANN, ANFIS, and SVM methods, this research demonstrates the model’s superior performance in handling monthly observational data. The findings underscore the importance of advanced modelling approaches in addressing complex hydrological phenomena, particularly in regions vulnerable to groundwater fluctuations. Ref. [82] study evaluates the effectiveness of the ELM for GWL prediction in Andhra Pradesh, India. Through a comparative analysis with other single and hybrid models like ANN, SVM, and GP, the research highlights the ELM model’s superior predictive capabilities, especially in Vizianagaram, India. This study underscores the significance of selecting appropriate modelling methods tailored to specific hydrogeological conditions. Ref. [83] investigate the effectiveness of the MLP-WA (Whale Algorithm) for predicting GWL in Yazd Province, Iran. This research compares the performance of MLP-WA with other hybrid models, such as RBF-WA and GP, focusing on temporal precipitation data as inputs. The findings highlight MLP-WA as the most effective hybrid model, outperforming other alternatives. This research emphasises the importance of incorporating wavelet analysis into machine learning methods to enhance predictive accuracy in GWL forecasting, particularly in regions with varying climatic conditions. Ref. [84] introduce the Hybrid Emotional Artificial Neural Network coupled with the Genetic Algorithm EANN-GA for estimating GWL in the Konan Basin, Japan. By combining FFNN with the Hybrid GTWANN approach, this study demonstrates robustness and effectiveness in providing accurate predictions, particularly in alluvial aquifers. Ref. [85] investigated the performance of hybrid Artificial Bee Colony and Particle Swarm Optimisation (ABC-PSO) ANN models for GWL forecasting in Manipal, Karnataka, India. This research highlights the significance of integrating optimisation algorithms with ANNs for enhanced predictive capabilities. Ref. [75] evaluated the Hybrid EANN-GA method for simulating spatial-temporal GWL in the coastal aquifer of the Konan Basin, Japan. This study demonstrates superior predictive skills by comparing the performance of EANN-GA with other models like GRNN and FFNN. Ref. [77] introduce the ANFIS-PSO method for GWL prediction in the Pahang watershed, Malaysia. This research demonstrates the model’s effectiveness in providing reliable predictions by optimising six different lag attributes using PSO, GA, DE, and ANFI methods. The findings suggest that ANFIS-PSO outperforms other models in terms of prediction interval width and accuracy. Table S6 (Supplementary Materials) summarises the evaluated articles that used hybrid methods and draws a general conclusion.

4. A Brief Discussion of the Literature Review Assessment

In exploratory research, it is vital to extensively evaluate the literature review and provide a comprehensive discussion that can offer valuable insights to readers. After reviewing the existing literature on AI-based methods for GWL modelling, this section distils and deliberates various fundamental abstract points.

4.1. Artificial Neural Networks

Through a data-driven approach, recent research in GWL modelling using ANNs has increased accuracy and reliability in predictions. Various studies indicate that ANN methods, such as FFANN, RNN, and ELM, perform exceptionally well. Several investigations demonstrate that ANNs exhibit higher accuracy than traditional statistical models like ARIMA due to their high capability to identify complex patterns and nonlinear relationships between input and output variables. Multiple studies have shown that the type and number of input variables significantly impact the model’s accuracy. For example, [18] demonstrated that combining water quality data with GWL significantly improves prediction accuracy. Similarly, [8] illustrated that using monthly evaporation and precipitation data as inputs for ANNs leads to a more accurate prediction of GWL. Different algorithms have been utilised to train ANNs, such as LM, BR, and BP [17]. These results highlight the crucial role of selecting the appropriate algorithm for training ANNs in improving prediction accuracy. Furthermore, hybrid methods, such as combining ANNs with ARIMA and optimising algorithms, have also improved prediction accuracy. For example, [13] demonstrated that learning machine methods are the best way to predict GWL, compared to more conventional ones. Incorporating geographical and hydrological data can help improve model accuracy. For instance, research by [12] demonstrated that using these data as ANN inputs improves daily GWL prediction accuracy. [21] successfully used the FFANN for hourly GWL prediction. A study by [86] also demonstrated that the BP-ANN performs better than the Integrated Time Series (ITS) in predicting GWL. Similarly, in research by [87], the RNN achieved higher accuracy in daily GWL prediction than the SVR method.

4.2. Adaptive Neuro-Fuzzy Inference Systems

Through comprehensive analysis of various articles in the field of GWL prediction, it is evident that the ANFIS method outperforms others, such as ANNs, SVMs, and GP. For instance, [34] reported the high accuracy of the ISA-LSSVM, a type of ANFIS, in predicting GWL. Indeed, ANFISs are capable of accepting and utilising diverse geo-environmental data. In [41], the NF-GMDH-PSO successfully predicted GWL using precipitation and temperature data. Comparisons with other methods often show that ANFISs have higher accuracy. Ref. [42] demonstrated that ANFISs outperformed SCFL FL in predicting GWL, achieving higher accuracy. While ANFISs are recognised as very effective for predictive modelling, there is still potential for improvement and optimisation. Enhancing model parameters, optimising training methods, and utilising larger datasets can increase accuracy. ANFISs are powerful tools in GWL modelling, renowned for its accuracy, ability to accommodate diverse data, and capability than other popular AI-based methods.

4.3. Support Vector Machines

Some factors contributing to the superior performance of SVM methods in predicting GWL can be mentioned:

• Handling complex data—SVMs are suitable for modelling complex relationships between variables due to its ability to separate complex and nonlinear data. For instance, in [55], SVM-ALO successfully predicted GWL;
• Robustness in heterogeneous data—data with different natures, such as temporal, environmental, and geographical variables, may be heterogeneous. SVMs perform well in predicting GWL by handling heterogeneous data effectively. For example, in [50], an SVM was successfully utilised to predict GWL;
• Performance in incomplete data conditions—water resource data often contain incomplete or outlier information. SVM’s resistance to outliers and ability to handle incomplete data perform well in real-world conditions. For instance, in [56], an SVM was successfully used to predict GWL.

4.4. Deep Learning

Recent advancements in DL have significantly improved the accuracy and reliability of GWL modelling. A synthesis of twelve prominent studies highlights the diverse methodologies, data inputs, and outcomes:

• Integrating advanced methods—XGBoost, LSTM, and CNN have revolutionised GWL modelling. For instance, [5] demonstrated the superior performance of XGBoost using a blend of time series, an ensemble-based method, and DL. This ability to integrate meteorological and hydrological datasets resulted in highly accurate 3-month forecasts, showcasing the potential of ensemble-based approaches;
• Innovative model combinations—combining different machine learning architectures with DL, like the SATCN-LSTM proposed by [57], enhances predictive accuracy. Their study utilised self-attention mechanisms alongside LSTM networks to model complex meteorological data, effectively outperforming traditional methods. This indicates that hybrid methods can better capture intricate data patterns;
• Long-term data utilisation—[6] leveraged extensive historical piezometric data (50 years) to reconstruct GWL fluctuations;
• Context-specific applications—[58,59] applied LSTM and SATCN-LSTM to address regional rainfall patterns and groundwater withdrawal scenarios, respectively;
• Handling outliers and noise—[60] tackled the challenge of data anomalies by developing a robust cost function integrated with recurrent ANN methods. This approach effectively managed outliers and noise, enhancing the reliability of GWL modelling. Such methods are crucial for real-world applications where data integrity can be compromised;
• Daily and seasonal predictions—the ability to provide accurate daily and seasonal predictions is exemplified by studies like those of [61,62]. The first one focused on daily GWL using LSTM with daily precipitation data, while the second recommended DL over ELM and GPR for seasonal groundwater depth predictions;
• Extreme weather event predictions—[63] emphasised accurately predicting groundwater responses to storm events using LSTM. Their study demonstrated that LSTM outperformed traditional RNN in such scenarios, providing critical information for flood management and coastal protection;
• Enhancing model performance with preprocessing—[66] highlighted the significance of comprehensive data preprocessing in improving model performance. Their use of NARX-DNN, LSTM, GRU, and ARX methods on water table level data showed that preprocessing steps are essential for accurately detecting stressed aquifer conditions;
• Superior methods comparisons—comparative studies, such as that of [67], evaluated different methods to identify the most effective ones for predicting water table depth. Their study found that their proposed model outperformed FFNN and Double-LSTM, highlighting the importance of rigorous model comparison in achieving optimal results.

Advanced and hybrid methods, long-term and context-specific data, robust preprocessing techniques, and tailored applications to regional and temporal conditions have significantly enhanced the accuracy and reliability of GWL predictive modelling.

4.5. Genetic Programming

GP methods have also been recognised recently as one of the more advanced tools for GWL predictive modelling. With its high flexibility and accuracy, this method enables the modelling of complex and nonlinear relationships in groundwater data. Research shows that GP and GEP are powerful tools. Combining these with optimisation methods and uncertainty analysis can enhance the accuracy and reliability of predictions. Some key findings that highlight the strengths and challenges of using GP in GWL modelling are:

• GEP and precise modelling—[4] has shown that the GEP, with its three-gene structure, is highly effective for GWL modelling. This method can accurately capture the inherent variations and complexities in the groundwater data using a time series dataset, providing precise and reliable predictions;
• Importance of weather station locations—[2] demonstrate that the precise placement of weather stations significantly impacts the accuracy of GP predictions. This finding is significant for optimising weather station networks;
• Comparison of GP with other methods—various studies, such as the research by [70], have shown that GPR outperforms GP in predicting groundwater salinity. Ref. [72] also revealed that GP performs better than the ANFIS in monthly GWL predictive modelling;
• Combining GP with optimisation and simulation methods—[73] highlight the high efficiency of combining GP with Continuous Ant Colony Optimisation and Monte Carlo Simulation;
• Uncertainty analysis in GP methods—[71] emphasise the importance of integrating uncertainty analysis into GP-based models. This integration improves the reliability and robustness of the models;
• Application of GP in identifying pollution sources—[68] show that the GP-based method for Bayesian experimental design is highly effective in identifying groundwater contaminant sources. This method, using MCMC, enhances the accuracy and efficiency of environmental monitoring efforts;
• Seasonal models—[74] indicates that linear regression approaches are suitable for predicting summer GWL, whereas ANNs are more effective for winter predictions.

4.6. Hybrid Algorithms

Hybrid Algorithms in GWL modelling represent a sophisticated convergence of multiple approaches and optimisation strategies tailored to address the intricacies of hydrological and hydrogeological systems. Follows a more specialised analysis, elucidating the intricacies and advancements in hybrid methodologies, supplemented with a concrete example:

• Utilising diverse methods—machine learning-based approaches, such as ANNs, SVMs, and MLP, are among the most prominent methods for predicting GWL. For example, [76] utilised the MLP–2NSGA-II–MOPSO to predict GWL;
• Hybridisation—combining various methods and employing Hybrid Algorithms to improve prediction accuracy is highly effective. For instance, [3] used the WT-LSSVM and Hybrid Algorithms to achieve more accurate GWL modelling;
• Application of optimisation algorithms—using algorithms such as genetic algorithms and evolutionary strategies to tune model parameters and improve prediction accuracy is highly effective. For example [1], utilised the PSO-ANN method to predict GWL;
• Integrating different data sources—incorporating diverse data sources such as weather and geological data to enhance prediction accuracy is crucial. For example, [82] utilised weather data and sea level information to model GWL;
• Utilisation of hybridisation and model fusion methods—many articles employ hybridisation to combine different approaches, such as ANNs with SVMs or ANNs with physical models like numerical flow models. For example, [80] utilised the AAEO-MLP to predict GWL;
• Incorporating uncertainty analysis—several studies emphasised the importance of uncertainty analysis in GWL modelling. For instance, [79] conducted a study comparing the ORELM with other hybrid methods and found it to be more accurate, efficient, and cost-effective. This underscores the significance of considering uncertainties in model predictions;
• Regional adaptation and validation—many articles focus on specific regions, considering local hydrogeological conditions and characteristics. For example, [83] developed the MLP-WA method to predict GWL, demonstrating its superiority over other hybrid approaches;
• Enhance model performance with novel methods—some studies introduce novel methods to enhance model performance. For instance, [84] proposed the Hybrid GTWANN, which demonstrated robustness and effectiveness in estimating GWL;
• Evaluation of model performance—it is crucial to assess the performance of prediction models rigorously. Studies such as those by [27,85] compared different models and found that hybrid approaches often outperform standalone methods regarding accuracy and reliability;
• Application of advanced methods for spatiotemporal prediction—several articles focus on the spatiotemporal prediction of GWL using advanced techniques like wavelet analysis and AI. For instance, [88,89] applied hybrid wavelet-AI methods to predict GWL.

5. Management Implications

Applying AI-based methods for GWL predictive modelling has profound implications for water resource planning, management, and decision-making policies. The insights derived from these advanced methodologies offer significant potential for improving the sustainability and efficiency of groundwater resources. The enhanced predictive accuracy is crucial for developing proactive groundwater management strategies, particularly in regions facing severe water scarcity and fluctuating water tables. Implementing the discussed AI-based methods facilitates the integration of real-time data and long-term predictions into groundwater management frameworks. This capability allows for more dynamic and adaptive management practises, ensuring that water extraction rates are aligned with the current and forecasted groundwater availability, thereby preventing over-extraction and safeguarding sustainable use [69,74].

Machine learning methods, such as ANNs and SVMs, in GWL studies can enhance cartographic models, improving their predictive accuracy and supporting more effective environmental strategies. The ability of the ANFIS method to handle uncertainty and capture complex relationships makes it essential for addressing the challenges associated with groundwater management and conservation [41]. For example, DL’s effectiveness in addressing hydrological forecasting challenges is crucial for optimising weather station networks and improving groundwater management practises, especially in arid and semi-arid regions where water resource sustainability is critical [73]. This last method is beneficial for areas facing significant groundwater contamination issues, as it helps in effectively identifying and mitigating sources of pollution. From the strategic placement of weather stations to uncertainty analysis, GP methods also play a significant role in developing sustainable groundwater management practises. Furthermore, Hybrid Algorithms that combine AI-based methods with traditional statistics have shown promise in areas with limited environmental data. These hybrid approaches enhance the robustness and reliability of GWL modelling, making them valuable tools for decision-makers in developing targeted interventions and policies that are resilient to data scarcity and environmental variability [1,75,76]. For example, [79] advocate adopting Hybrid Algorithms to improve groundwater management strategies in diverse hydrogeological settings. Integrating optimisation algorithms with ANN, ANFIS, and SVM methods provides valuable insights into improving groundwater management strategies by adopting more robust and reliable hybrid modelling methods [85].

AI-based management methods can help develop warning systems for droughts and groundwater depletion. These systems provide accurate predictions and timely alerts to authorities and communities about impending water shortages. This allows for the implementation of mitigation measures well in advance [68,71]. Additionally, AI-based models can optimise the allocation of groundwater resources for various purposes, such as agricultural irrigation, industrial processes, and domestic consumption. By simulating different usage scenarios, policymakers can assess the potential impacts of varying allocation strategies and choose the most efficient and fair distribution plans [72,77].

Integrating AI-based GWL modelling into water resource planning and management practises offers numerous benefits, including improved prediction accuracy, enhanced decision-making capabilities, and the development of more sustainable strategies. As these methods advance, their applications in groundwater management are expected to become increasingly sophisticated, providing crucial support for addressing the complex challenges of water resource sustainability. These advancements offer essential tools to water resource managers, policymakers, and environmental planners, enabling them to make informed decisions to ensure sustainable groundwater management in the face of changing climatic and ecological conditions.

6. Future Works

Based on the provided recommendations for future studies in various fields of GWL modelling, some crucial suggestions include:

• Integrating data and methods—combining diverse data sources and AI-based methods can enhance prediction accuracy;
• Algorithm optimisation—optimising learning algorithms and employing advanced algorithms like ELM and SVR can improve model accuracy;
• Geographical expansion—investigations across different geographical locations and varying weather conditions can aid in model generalisation;
• Interdisciplinary studies—employing interdisciplinary methods integrating data from various scientific fields can enhance model accuracy and efficiency;
• Precise variable selection and ANFIS parameter optimisation—precise variable selection and model parameter optimisation are crucial for using ANFISs in GWL modelling;
• Expansion of uncertainty analysis—utilising advanced methods for uncertainty analysis in GP-based methods can enhance prediction reliability, especially in highly uncertain scenarios;
• Development of real-time prediction systems—creating real-time prediction systems capable of processing data quickly can significantly aid in water resource management;
• Long-term model evaluation—evaluating model performance over extended periods can identify strengths and weaknesses, guiding further improvements;
• Considering climate change effects—developing models that account for climate change’s impacts on GWL and distribution is essential.

7. Final Remarks

The performed comprehensive literature review, covering the time interval 2001–2023, highlights the use of AI-based methods in GWL modelling. The examination of the different studies shows several critical results for the most used methodological approaches, as follows:

• ANNs and other advanced algorithms, such as FFANN, RNN, and ELM, have significantly improved the accuracy and reliability of GWL modelling. Compared to traditional statistical models, these methods better capture complex patterns and nonlinear relationships between input and output variables. In addition, including different input variables and optimising the algorithms contribute to improving the prediction accuracy;
• ANFISs are robust methods for GWL prediction and outperform others such as ANNs, SVMs and GP in various analysed papers. ANFISs can account for different environmental data and provide higher accuracy than alternative methods. Combining fuzzy logic with neural networks shows higher accuracy, especially for non-stationary time series data. The choice of membership functions significantly impacts the accuracy of the constructed GWL prediction models;
• SVMs are superior in predicting GWL as they can effectively handle complex and heterogeneous data. This method is characterised by its robustness in dealing with incomplete data and outliers. SVMs, especially those using the RBF kernel algorithm, are excellent at capturing the complexity of GWL;
• Recent advances in DL methods, including XGBoost, LSTM, and CNN, have significantly improved the accuracy and reliability of GWL modelling. These methods are characterised by their ability to integrate different data sources, provide accurate forecasts for different temporal and spatial scales, and cope with extreme weather events;
• GP and GEP methods are powerful tools for groundwater table modelling. GP is characterised by high flexibility and accuracy and allows the modelling of complex and nonlinear relationships in groundwater data. In addition, including optimisation methods and uncertainty analyses increases the reliability and robustness of the predictions;
• Hybrid methods combine different modelling approaches and optimisation strategies and have a higher forecasting capability. These methods utilise the strengths of different approaches and incorporate different data sources to improve forecast accuracy and reliability. In addition, including uncertainty analyses and adapting models to specific local and regional environmental conditions help advance GWL modelling.

The preceding literature review emphasises the transformative impact of AI-based methods and Hybrid Algorithms on GWL modelling. These advancements offer valuable insights and tools for effective groundwater planning, management, and sustainable water resource utilisation in various environmental and territorial contexts. AI-based methods for GWL modelling contribute to improved predictive accuracy, support adaptive and sustainable management practises, particularly in arid and semi-arid regions, and play a crucial role in addressing challenges related to groundwater sustainability in the face of climate change and increasing water demands. Continued research and investment in AI are vital to ensure effective and sustainable management of essential groundwater resources.