A Bibliometric-Systematic Literature Review (B-SLR) of Machine Learning-Based Water Quality Prediction: Current State and Projections

1. Introduction

Surface and groundwater pollution is a problem of global concern, as it can negatively affect freshwater availability, a key aspect for the conservation of ecosystems and the health of the population. This is aggravated by the strong pressure exerted by a growing demand for fresh water to sustain population growth and associated economic development [1,2,3]. Thus, the availability of good-quality freshwater sources has a direct impact on the ability to ensure water security, especially when nearly 2.2 billion people lack access to safe drinking water and another 4.2 billion do not have safe sanitation systems [4].

Water quality is determined by different factors, such as weather conditions [5], seasonal changes in hydrological regime [6], local geological conditions [7], human activities or land use, among others. Water quality has traditionally been assessed with in situ probes or by sampling and subsequent analysis of biological and physical-chemical parameters in the laboratory. Finally, the water quality has been increasingly forecasted with physically-based simulation models, such as QUAL2K [8,9,10] and WASP [11,12], as well as statistical models such as multivariate analysis [13,14], and linear regression [15,16].

In this context, the application in recent decades of artificial intelligence (AI) [17,18,19,20], especially machine learning (ML) models [21,22,23,24] and deep learning (DL) [25,26,27,28,29,30,31], have contributed significantly to the development of water quality prediction techniques for environmental management purposes, strengthening the path towards the fulfillment of the water-related Sustainable Development Goals (SDGs) [32]. These prediction models can handle large amounts of data and adapt to complex relationships. In fact, the use of ML for water quality studies is highlighted as an example of the United Nations Global Acceleration Framework, for the scope of the SDG-6 [33,34].

An important aspect in the application of AI/ML/DL for water quality prediction is that its use allows anticipating variations in physical-chemical and biological parameters of water, using either historical data, remote sensing, or hydrological models, therefore optimizing environmental management and evidence-based decision-making. In this regard, two main approaches are identified in the applications used for water quality prediction: a) those that seek the prediction of individual parameters known as water quality parameters (WQP) [10,35,36,37,38] and (b) those whose purpose is to integrate these parameters into a water quality indicator, traditionally known as Water Quality Index (WQI) [39,40,41,42,43]. Regarding the former, we can mention, only as a given example, the prediction of total phosphorus concentrations in Taihu Lake, China, using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm and the Long Short-Term Memory (LSTM) algorithm together with the use of Shapley Additive Explanations (SHAP) as an interpretability method [44]. Another example is the study [45], which applies the Random Forest (RF) method for the prediction of 14 physicochemical parameters using historical data from the Loa River, in the xeric region of Chile. With respect to the second approach, an example is the study [46], where a set of ML models is used to estimate the WQI in the Bug River, Ukraine. A high predictive performance was also found in the study [47] where different ML models (i.e., Extreme Gradient Boosting (XGB), Support Vector Regression (SVR), Artificial Neural Network (ANN), Random Forest Regression (RFR)) were evaluated to predict the Entropy Weighted Water Quality Index (EWQI) and estimate groundwater quality in Kerala, India.

The aforementioned studies are examples of how AI/ML/DL methods have been consolidated as water quality prediction techniques, given their robustness and efficiency to extract patterns of complex systems, as water systems, making these techniques an interesting alternative to conventional water quality modeling techniques [48,49,50]. Indeed, systematic reviews of literature at a global level show the high interest of researchers in the use of AI/ML/DL for water quality prediction. For example, the authors in [51] reviewed 876 articles within the period 2015-2022, showing that the United States, England, Iran, India, and China have emerged as major contributors to the field of water quality prediction with ML and DL. Likewise, the authors in [52] reviewed 249 articles on water quality using Internet of Things (IoT) models and machine learning. In another study [53], 253 articles were reviewed, finding that DL optimizes the processing of large volumes of data through parallel computing, facilitating the effective prediction of water quality, although its success depends on the quality of the data used. Finally, in [54], the literature was reviewed in terms of groundwater quality prediction studies using AI/ML/DL, finding that the ANN, Adaptive Neuro-Fuzzy Inference System (ANFIS) and Support Vector Machine (SVM) techniques have proven to be efficient and accurate tools for such purposes.

Despite the intrinsic value of the aforementioned literature review studies, it is important to highlight a key conceptual distinction between the two main review approaches. One is known as Bibliometric Analysis, that is, a quantitative method that allows the identification of emerging trends, thematic networks, and research gaps [55,56,57]. Another corresponds to Systematic Review of the literature, a rigorous and structured method used to collect and evaluate all available evidence on a research question [58,59,60]. Although the two approaches complement each other, and together they are a promising methodology to elaborate robust reviews of the scientific literature, to date no work, to the best of our knowledge, has addressed the study of the state of the art on the application of AI/ML/DL for water quality prediction through the integration of both approaches. Thus, such a dual-approach review corresponds to the main objective of this work, as this contribution proposes to analyze the state of the art on the application of AI/ML/DL for water quality prediction in the 2000-2024 time frame, using the Bibliometric-Systematic approach Literature Review (B-SLR) [61]. It is hoped that the results of this work can guide researchers interested in the study, assessment and prediction of water quality using AI/ML/DL techniques.

2. Materials and Methods

We employed a structured framework for conducting systematic literature reviews, known as the Bibliometric-Systematic Literature Review (B-SLR), following the guidelines outlined in [61]. This approach was used to assess the current status, emerging trends, and research gaps in the application of AI/ML/DL techniques to freshwater (surface and groundwater) quality studies. The methodology B-SLR combines the bibliometric analysis (BA) [57,62,63] with the systematic literature review (SLR) [64,65,66]. Thus, the B-SLR approach facilitates the broadening of topic scope, expanding the domain of knowledge available to researchers working in the field [60]. According to the adopted methodology, this work was developed in three sequential stages: a) Data collection, b) Bibliometric analysis, and c) Systematic Review of the Literature, as illustrated in Figure 1.

2.1. Data Gathering Process

This stage aims to identify the research topic and the structural framework to establish the research questions, keywords, search strings, database, time scope of the research, techniques to carry out the search and analysis tools, where each stage generates a result that serves as an input for the next stage.

The research questions raised in this study seek to identify a set of patterns behind the application of AI/ML/DL techniques in water quality prediction. Table 1 presents the guiding research questions of this study with their corresponding justification, following the guidelines of the systematic review of the literature [65].

To obtain the bibliographic information necessary to answer the research questions posed, the following line of code was generated for the search and construction of the collection of bibliographic references:

water AND quality AND prediction AND machine AND learning) OR TITLE-ABS-KEY (water AND quality AND prediction AND artificial AND intelligence) OR TITLE-ABS-KEY (water AND quality AND prediction AND deep AND learning))

AND PUBYEAR > 1999 AND PUBYEAR < 2025 AND (LIMIT-TO (DOCTYPE, "re”) OR LIMIT-TO (DOCTYPE, "ar”)) AND (LIMIT-TO (LANGUAGE, "English”)) AND (LIMIT-TO (SUBJAREA, "ENVI”) OR LIMIT-TO (SUBJAREA, "ENGI”) OR LIMIT-TO (SUBJAREA, "EART”) OR LIMIT-TO (SUBJAREA, "MULT”)).

The temporal scope of this study was defined from January 2000 to December 2024. Bibliographic references were obtained from the Scopus database (https://www.scopus.com/home.uri), which offers a comprehensive and global overview of scientific production. Scopus is one of the most widely used sources in bibliometric analysis and is recognized for its reliability and high-quality data, particularly in academic research related to water resources and water quality [17,56,57,67,68,69].

The selection criteria for documents included in the bibliographic reference collection were as follows: a) Publications classified as “Research Article” or “Review,” excluding conference proceedings, books, book chapters, theses, and reports; b) Publications written in English; c) Publications within the subject areas of Environmental Science, Engineering, Earth and Planetary Sciences, and Multidisciplinary; d) Articles containing the keywords specified in the search string. In this first retrieval, a total of 1822 articles were obtained.

Bibliographic references were exported from Scopus in CSV format and subsequently imported into the R programming environment [63]. A comprehensive list of 29,144 Scimago journals (https://www.scimagojr.com/) for the year 2023 was incorporated to filter the 1,822 original articles selected for B-SLR analysis, retaining only those published in journals classified within Q1 and Q2 quartiles, in accordance with the recommendations outlined in reference [61]. Publications to 2014 were excluded, establishing the analysis period between 2015 and 2024. This procedure involved excluding publications prior to 2014, thereby establishing the analysis time frame from 2015 to 2024. In addition, a systematic cleaning process was conducted, which included the removal of duplicate rows, incomplete records in the fields Index Keywords, Author Keywords, DOI, and Abstract, as well as the elimination of entries with duplicated DOIs and false positives [70], that is. These publications did not align with the objective of the present study.

The automation of this process was achieved through the implementation of an iterative text mining approach based on Topic Modeling, using the TextMiner package in R as a specialized tool for constructing and analyzing thematic models [71]. The procedure consisted of: a) Merging the Title, Abstract, and Author Keywords fields from Scopus into a single column; b) Applying the Topic Modeling procedure to detect reference topics for each record; c) Grouping the records into 30 clusters, each characterized by the five most frequent topics; and d) Identifying terms that, being representative of the Title, Keywords, and Abstract, indicated that the publication fell outside the scope of the study.

Records containing the terms selected in step d) were removed from the database, and a new iteration was performed. In each iteration, the appropriateness of each removal was confirmed through manual inspection.

The record elimination procedure, assisted by Topic Modeling, resulted in a preliminary database of 827 records. To ensure that the publications were aligned with the scope of the study, only those published in journals whose titles included the expressions Hydro/ Enviro/Water were retained. Furthermore, only journals with at least three published articles were considered. This filtering process yielded a refined database of 424 articles. Subsequently, all records underwent manual inspection, which involved reviewing the title, abstract, introduction, methodology, and conclusions to exclude publications that did not meet the thematic criteria of the study. After this final validation step, the resulting database comprised 276 articles published between 2015 and 2024. These records were used in the subsequent stages of the bibliometric analysis (BA) and systematic literature review (SLR).

2.2. Bibliometric Analysis (BA)

The bibliometric analysis was conducted using the Bibliometric package in the R programming language [63]. This package has been widely applied in bibliometric studies related to hydrology [56,72] and water quality [73,74]. The analysis included the generation and visualization of maps and graphs corresponding to performance analysis [17,56,62] and science mapping [57,75,76].

2.3. Systematic Literature Review (SLR)

A comprehensive reading stage was conducted for the 276 selected articles, conducted using the Systematic Literature Review (SLR) approach described in [60], in order to extract key data and information to address the research questions. This methodology enabled the identification, collection, analysis, and synthesis of relevant information to answer the research questions posed [77]. Its robustness stems primarily from the transparency of its implementation process, which ensures the reproducibility of the review[78,79].

Each study was classified as either a Review or a Research article. Within the Research category, articles were further classified according to the type of water body under study: surface water or groundwater. A particular case involved two Review articles [80,81] that addressed both systems (groundwater and surface water); this condition did not affect the classification by water body type, as these studies were not part of the Research category.

3. Results

3.1. Bibliometric Performance Analysis

The annual scientific output on water quality prediction using AI/ML/DL has increased in the study period, as shown in Figure 2, exhibiting an exponential increase, especially marked since 2020.

The growing interest in research on the prediction of water quality using AI/ML/DL is within a global context driven by climate change, and increased pollution and demand for fresh water [5,82]. This evolution indicates that the analyzed domain has passed an initial stage of exploration, positioning itself in a phase of maturity. Likewise, it is possible to expect an even more significant development in the coming years, anticipating a continuous strengthening of both interest and research in this field [83,84].

3.1.1. Journals

When ranking the top 10 of the 32 sources based on the H-index and Total Citations (TC) [85], it is observed that the journals Water (Switzerland), Journal of Hydrology, Environmental Science and Pollution Research, Science of the Total Environment and Water Research are the publications with the highest number of papers (first 5) within the reference collection (Table 2).

3.1.2. Most Cited Documents

The most cited documents within the analyzed collection of bibliographic references were identified. Both global and local citation counts were considered to assess the reach and influence of the research from complementary perspectives. Local citations reflect the impact of a document within the specific dataset under analysis, while global citations represent the total number of times each article has been cited in the Scopus database.

Table 3 provides the top 10 most cited publications and corresponding (leading) authors in terms of local and global citations. The referred papers reveal important contributions to research on water quality prediction using AI/ML/DL techniques.

According to Table 3, the most cited reference corresponds to the work of Barzegar [86] entitled "Short-term water quality variable prediction using a hybrid CNN–LSTM deep learning model", which investigates the prediction of water quality in Lake Prespa, Greece, and proposes a hybrid CNN-LSTM model to predict dissolved oxygen (DO; mg/L) and chlorophyll-a (Chl-a; μg/L).

Second, the study [87] compares different machine learning methods (such as artificial neural networks and support vector machines) for the prediction of water quality parameters of northwest Iran’s Aji-Chay River, demonstrating the effectiveness of these techniques in improving environmental monitoring. Its high number of global citations reflects the interest in using water quality prediction to support the transition toward more sustainable water resources management.

Third, the work of [88] examined the application of a decision tree model for predicting the Water Quality Index (WQI) in the Klang River. The research successfully showed that the number of water quality parameters needed for monitoring can be reduced while maintaining prediction accuracy above a 75% benchmark. Finally, the most recent study of the list [93], proposes a hybrid approach that combines the Random Forest algorithm with an improved version of the SMO algorithm for support vector machines. Applied to the Saf-Saf River Basin, this model improved WQI prediction, highlighting the growing role of optimized hybrid approaches in watershed management.

3.2. Bibliometric Science Mapping

3.2.1. Network Analysis on Co-Occurrence of Authors’ Keywords

The network analysis of co-occurrence of author keywords presented in Figure 3 reveals a conceptual structure derived from studies on machine learning and water quality management. Each node represents a keyword, with its size indicating the degree of connectivity (number of links to other terms). Connecting lines reflect co-occurrence relationships, and their thickness represents the strength of those associations. Nodes are grouped into three thematic communities, distinguished by color: group 1 (blue), group 2 (orange), and group 3 (green), enabling the identification of conceptual subdomains within the field.

Central terms such as support vector machine, deep learning, water quality predictions, and water quality management, along with technical acronyms like ML (machine learning), AI (artificial intelligence), SHAP (SHapley Additive exPlanations), ANN, CNN (convolutional neural networks), LSTM (long short-term memory), and random forest, exhibit high connectivity. This suggests their integrative role in the analyzed literature and highlights their methodological relevance in the development of predictive and explanatory models applied to water systems. The visualization supports the identification of thematic clusters, methodological relationships, and emerging areas at the intersection of artificial intelligence and environmental management.

3.2.2. World Cloud

Another method for identifying the frequency of the most common terms used by the authors in the collection corresponds to the word cloud generated from the keywords. Figure 4 shows a predominant interest in “water quality,” which serves as the central axis of the study corpus. Its association with indicators such as “WQI,” “WQP,” and water bodies like “groundwater quality” and “surface water” reaffirms the environmental orientation of the studies. From a methodological standpoint, there is a predominant interest in techniques such as “random forest,” “deep learning,” “ML,” and “ANN,” along with specific architectures like “LSTM,” “CNN,” “RNN,” and “SVM,” used in modeling and prediction processes of hydrological parameters.

Complementing this trend is the appearance of terms like “hybrid model,” “ensemble learning,” and “transfer learning,” reflecting the adoption of integrated strategies and techniques for transferring pretrained representations aimed at improving model generalization and efficiency. Finally, “XAI” and “SHAP” reveal an interest in model explainability, pointing toward advanced analytical systems for the management and evaluation of water sustainability.

3.2.3. Thematic Map with Authors’ Keywords

Figure 5 presents a thematic map derived from bibliometric analysis, grouping clusters associated with distinct research subtopics in the field of water quality prediction using AI, ML, and DL. This map enables the evaluation of each topic in terms of centrality and maturity, identifying established, emerging, and declining areas. In the quadrant of Basic themes, terms like “water quality prediction,” “deep learning,” and “LSTM” reflect the consolidation of deep neural networks as a methodological foundation. Notably, LSTM networks have been highlighted in multiple studies for their strength in modeling complex temporal dynamics in water quality parameter prediction [31,86,94,95]. Furthermore, hybrid models and the ANFIS technique have provided enhanced precision and flexibility in complex scenarios, particularly in surface water contexts [87,96,97,98].

In the quadrant of Motor themes, terms such as “WQI,” “ANN,” “AI,” and “ML” stand out, indicating that artificial neural networks and machine learning models have been extensively studied and applied, yielding significant results. Conversely, “transfer learning” emerges as a nascent exploratory line within water quality prediction, although several studies [6,94,99,100,101,102] have demonstrated its potential to improve predictions in data-scarce environments.

In the quadrant of Niche themes, the terms “data-driven models,” “lake water quality,” and “prediction intervals” suggest that lake water analysis and data-driven modeling represent growing niches within water-related research, though they have yet to become dominant in AI/ML/DL-based water quality prediction.

In the quadrant of Emerging or declining themes, terms such as “support vector machine,” “feature selection,” and “gaussian process regression” indicate that these methods (SVR, GPR), although historically robust and widely used in this field, may be losing relevance compared to more modern approaches such as deep learning. Additionally, procedures like feature selection may have already transitioned into standard practice within AI/ML/DL applications. Meanwhile, “extreme gradient boosting” and “SHAP” are associated with advanced ML techniques but have not yet reached the prominence of core topics. The emerging discipline known as Explainable Artificial Intelligence (XAI), or Interpretable Machine Learning (iML), seeks to address the challenges posed by the opacity of black-box techniques in AI/ML/DL. Its application has expanded in recent years across various engineering domains [83,103,104,105], and has recently been implemented in the field of water quality, particularly in groundwater studies [106,107].

3.2.4. Thematic Evolution and Trend Topics with Keywords Plus

The thematic evolution in the BA is useful to visualize more clearly the change and evolution of research topics over time, allowing the identification of windows of opportunity for future research work. To have a more global view of the thematic evolution and current trends in water quality prediction using AI/ML/DL, Figure 6 presents the thematic evolution based on the Keywords Plus. These words are automatically generated by the BA algorithms from sources such as the titles of the references cited in the documents, allowing for to assessment, from a complementary perspective, of the areas that are related to the central axis of the research search. Thus, the application of models based on ANN and decision support techniques is maintained over time, while models based on SVM have been less constant in recent years.

In recent years, cutting-edge ML and DL techniques have increasingly been applied to environmental challenges such as pollution, climate change, water quality, and water availability. These approaches are often integrated with Geographic Information Systems (GIS) technologies [1,2,108,109,110], enhancing spatial analysis capabilities.

While GIS is now widely used in the field of water resources, it has yet to become a dominant tool for predicting water quality through AI/ML/DL methods. Notably, interest in developing predictive models for surface water quality has remained consistent throughout the study period. Since 2015, ML has emerged as the leading approach, demonstrating strong applicability in environmental monitoring, wastewater treatment, and water quality assessment, indicating a sustained research focus on water resources management.

3.2.5. Social Structure

Figure 7 presents the map of main collaborations between countries, which reveals a highly interconnected structure of scientific cooperation. Countries such as China and the United States stand out, acting as central nodes, evidencing their leadership in the production of knowledge and in the articulation of international efforts. This dynamic reflects not only the shared interest in addressing the challenges of water quality prediction but also the growing need to integrate data, methodologies, and interdisciplinary approaches on a global scale. Countries such as India, Australia, the United Kingdom, Italy, and Germany stand out for their collaborative connections as well.

However, the map also reveals a striking absence of Latin American countries, suggesting a potential asymmetry in scientific visibility. This underrepresentation may be partially attributed to the methodological filter applied in this study, which included only publications indexed in high-impact Q1 and Q2 journals. While this criterion ensures scientific rigor, it may inadvertently exclude valuable regional research that, due to structural or editorial limitations, does not reach these publication venues. As a result, the map not only visualizes collaboration intensity but also reflects broader epistemic inequalities in global scientific discourse.

3.3. Systematic Literature Reviews Results

The systematic review allowed us to classify each work into different categories according to: Approach (Review or Research articles), body of water of the study, that is, Surface water (rivers, reservoirs and lakes) or Groundwater, as presented in Figure 8. Figure 8(a) shows the classification of the 274 articles collected according to their type: Research Articles and Review Articles. The disparity between the two approaches suggests that, although research in the area is growing rapidly, a significant gap persists in terms of consolidating knowledge through systematic reviews of the state of the art. This represents a valuable opportunity to strengthen the field through studies that integrate and synthesize existing findings (such as ours).

Within Research articles, 75% (n = 190) correspond to studies focused on surface water quality, while the remaining 25% (n = 65) address groundwater quality issues (Figure 8(b)). This bias towards the study of surface waters can be attributed to the relevance of watershed monitoring for the sustainable management of water resources. This monitoring allows us to understand the complex interactions between biological, chemical, physical, and environmental factors that determine water quality, as well as to anticipate the appearance of alterations to water quality. This information is crucial for informed decision-making and long-term planning, which are essential for preserving the health of water bodies and the ecosystems that depend on them [111].

In Figure 8(c), we present the distribution of works considering only surface water bodies, i.e., rivers, lakes, and reservoirs. It is shown a major focus in rivers, which could be because they are particularly exposed to pollution caused by anthropogenic activities [40,112]. In addition, during periods of drought, rivers are more easily and rapidly affected by the reduction of the flow, affecting the quality of the water in terms of its physical, chemical and biological properties [113]. The study [58], made a significant contribution with the exhaustive review of the literature in the period 2000-2020 on the modeling of river water quality in the world, reviewing 209 research articles from Scopus journals. The results showed that most of the study areas are Asian countries, such as China, Iran, India, Malaysia, Taiwan, Korea, Iraq, Bangladesh and Thailand, accounting for more than 50% of research papers. These results are consistent with the findings of the current work for the period 2015-2024. Based on the classification shown in Figure 8, this study delves into the two water bodies that are currently of greatest interest in the application of AI/ML/DL techniques for water quality prediction: rivers and groundwater. For this reason, a detailed description of the literature for them is presented below.

3.3.1. Prediction of River Water Quality Using AI/ML/DL

Focusing on the publications on water quality prediction in rivers, Table 4 presents the characterization of a random sample of the database, equivalent to 40% (n=57).

Recent advances in AI/ML/DL techniques have significantly enhanced the prediction of river water quality, as demonstrated by the 57 studies summarized in Table 4. These investigations employ a range of algorithmic approaches, including CNN-LSTM, XGB, RF, LSTM, SVM, and hybrid models to estimate key indicators such as the Water Quality Parameter (WQP) and the Water Quality Index (WQI) across rivers in Asia, America, Europe, and Africa. Artificial Neural Networks (ANNs) have played a prominent role in this domain, owing to their capacity to model complex nonlinear relationships and anticipate fluctuations in water quality, as confirmed by studies [133] and [138]. In [43], ANN-based models were evaluated alongside Gradient Boosted Trees (GBT), Decision Trees (DT), Support Vector Machines (SVM), and Random Forests (RF) for predicting the WQI of the Indian river system, with GBT achieving the highest performance. Similarly, [155] implemented four standalone decision tree models and twelve hybrid configurations to estimate the WQI of the Talar River in Iran. The Bagging-Random Tree (BA-RT) approach yielded the most accurate results, reinforcing the superiority of hybrid models over simpler alternatives, as also noted in [159].

The integration of optimization algorithms has proven to significantly enhance the accuracy of machine learning (ML) models in predicting river water quality. For instance, studies [25] and [93] employed Particle Swarm Optimization (PSO) and Sequential Minimal Optimization (SMO), respectively, to strengthen models such as PSO-DBN-LSSVR for the Juhe River in China and SMO-SVM for the Wadi Saf-Saf River in Algeria. Additionally, the application of Wavelet Transform (WT) has been key in identifying relevant variables and reducing noise, as demonstrated in studies on the Aji-Chay River (Iran) [87], Fujian (China) [101], Dongjiang (China) [145], and Sefid Rud (Iran) [154]. These enhancements have led to improved predictive performance in models such as WT-SVM, WT-RF, and WT-ANN. Another widely applicable technique is XGB, which has been successfully implemented in basins such as the Delaware River (USA) [115], the Han River (South Korea), and regions of the northwestern United States [140], due to its ability to capture complex interactions among predictors and deliver high predictive accuracy.

Hybrid architectures that integrate Convolutional Neural Networks (CNNs) with LSTM units have also demonstrated strong performance in river water quality prediction, as shown in studies on the Yangtze River [114,156]. In [112], the SSA-CNN-LSTM model was applied to the Sheshui River, achieving effective integration of temporal and spatial patterns for estimating the Water Quality Parameter (WQP). Other studies have explored variants such as Gated Recurrent Units (GRU) [122,125], Bidirectional GRU (BiGRU) networks [148], and hybrid approaches like ANFIS–GP and ANFIS–SC [143] to address nonlinear modeling challenges. In this context, emerging techniques such as Transfer Learning (TL) have shown considerable promise by enabling the reuse of previously acquired knowledge in new environments, thereby enhancing predictive performance as demonstrated in studies on the Fujian river system in China [100,101].

Finally, the geographical diversity of the studies listed in Table 4 spans a wide range of river systems, including the Yangtze [114,156], Sheshui [112], Tanjiang [121], Li and Liu [125], Pearl [135], Fuyang [136], Xiaofu [128], Lijiang [129], Juhe [25], Euphrates [126,143], Júcar [132], Yamuna [31,143], Langat [139], Klang [146], Sefid Rud [154], Talar [134], Sefidrood [92], Aji-Chay [87], Nakdong [147,153], Oyster River [124], Upper Red River Basin [100], Danube [142], Burnett [152], Kelantan [138], and Bullfrog [133]. These rivers span diverse climatic zones from tropical and temperate to arid and mountainous and reflect a broad spectrum of hydrological and geological contexts. This territorial breadth highlights the versatility of AI/ML/DL models in adapting to varied environmental conditions, reinforcing their value as effective tools for the sustainable management of water resources.

3.3.2. Prediction of River Water Quality Using AI/ML/DL

Regarding groundwater, Table 5 presents a random sample of 40% (n=26) of the publications on groundwater quality prediction found in our database, spanning various regions of the world. Diverse environmental, hydrogeological, socioeconomic, and technological factors have shaped the development of advanced computational approaches for predicting groundwater quality.

These approaches range from traditional statistical models and standalone machine learning algorithms (e.g., decision trees, support vector machines, etc.) to hybrid, ensemble methods and SHAP-enhanced models. They also include the integration of geospatial tools, remote sensing data, and cloud-based platforms for data processing and visualization. The most frequently studied contaminants are nitrates, arsenic, fluoride, heavy metals, and total dissolved solids. In addition to studies focused on individual water quality parameters (WQPs), there is research incorporating integrative water quality indices such as the Water Quality Index (WQI), Irrigation Water Quality Index (IWQI), Entropy-weighted Water Quality Index (EWQI), and Groundwater Quality Index (GWQI).

The wide variety of AI/ML/DL models highlights a transition from traditional approaches such as ANN, LSTM, Multilayer Perceptron (MLP) and CNN, to more sophisticated models such as XGB and LightGB. These models are generally applied in studies that require modeling nonlinear relationships or time series, and have been widely used to estimate water quality indices as well. From the reading of the studies listed in Table 5, it is possible to identify a current trend in the use of ensemble-type algorithms, such as Random Forest, Gradient Boosting and Bagging, which are considered robust methods capable of handling the high dimensionality and multicollinearity of predictive models [47,162].

Other ensemble algorithms such as CatBoost, Bagging, and Extra Trees have also been frequently used to predict specific WQP such as nitrates [160,163,174,182], salinity levels [162,171], metals [161,178], as well as water quality indices such as WQI, IWQI, EWQI, and GWQ. These algorithms allow multiple hydrochemical variables to be handled, and reliable predictions to be constructed in complex environments. Regression models have also been widely used in this area, such as the use of Multinomial Logistic Regression (MnLR), which allows water quality to be classified into multiple categories, being useful in environmental risk and zoning studies [176].

Among the emerging models, Generative Adversarial Network (GAN) [161] and Group Method of Data Handling (GMDH) [165] stand out. They are generally used in contexts with complex dynamics, particularly with scarce data, and be particularly useful in predictions of Sr²⁺ and salinity levels [161]. In addition, explanatory models such as SHAP and LIME represent a significant advance in the interpretability and explainability of predictive models applied to groundwater quality [178]. These techniques, typical of the field of XAI, are important because they allow the outputs of complex algorithms to be broken down into quantifiable contributions of each input variable, allowing one to understand how much each factor influences the final prediction. For example, in [162] SHAP was applied to assess the impact of physicochemical factors on salinity levels in multiple aquifers, revealing key spatial patterns using interpretable maps. Similarly, [163] used SHAP in a nitrate prediction model to identify the main geoenvironmental pollution-related drivers in a UK aquifer, integrating interpretation and prediction into a unified framework.

In this sense, the integration of GIS technologies, together with XAI models, strengthens the capacity to spatially represent the results, detect risk areas and generate useful tools for water management. Together, the AI/ML/DL models allow the prediction of groundwater quality to be approached from an interdisciplinary, explanatory and applicable perspective, configuring themselves as key tools for water sustainability in complex environmental scenarios [106,168].

On the other hand, in the hydrogeological context, studies are conducted in regions characterized by a wide range of factors, including aquifers affected by saline intrusion (Vietnam, Iran), arid zones (Algeria, Saudi Arabia), and areas with carbonate lithology that induce water hardness (India, USA). Natural processes such as weathering, evapotranspiration, and the influence of volcanic emissions also contribute to the spatial and temporal variability of hydrochemical parameters [161,177,178]. These hydrogeological factors, combined with local climatic conditions, shape physically complex and dynamic environments that demand predictive models with high adaptive capacity. In this context, the suitability of groundwater quality for human consumption is conditioned by these factors, among them, variations in well depth, which can significantly alter water mineralization and contaminant mobility due to interactions with lithological formations, hydraulic gradients, and specific redox conditions [172,184]. Therefore, the applied algorithms must be capable of capturing aquifer heterogeneity by integrating geological and hydrodynamic variables that directly influence groundwater quality.

Finally, at the regional level, Asia leads scientific production, with strong representation from India, China, Iran, and Bangladesh. These countries face severe water stress, which justifies the scientific community’s growing interest in groundwater quality and the development of advanced predictive models.

4. Answering the Research Questions

Consistent with the B-SLR framework adopted in this study, the driven research questions are answered below.

RQ1. What are the most widely used AI/ML/DL algorithms in water quality prediction?

The literature review showed that assembly models, such as Bagging and Boosting, have established themselves as predominant techniques in the prediction of both surface water and groundwater quality [185]. Their effectiveness is supported by studies that demonstrate their ability to reduce mean absolute errors and quadratic errors [186]. In addition, recent research has explored variants of these models, such as Grid Search Random Forest (GS-RF) and XGB, to optimize prediction accuracy on parameters such as turbidity and different nutrients [187], showing that algorithms such as CatBoost Regression (CBR) offer advantages in terms of stability and adaptability, particularly in handling heterogeneous datasets, minimizing overfitting, and maintaining consistent performance across varying environmental conditions and input configurations [171].

These assembly models have been successfully applied in diverse hydrological contexts, including urban rivers and Rural [58], and their use extends to the prediction of water quality indices in multivariable and complex scenarios [188]. Moreover, their ability to handle large volumes of data and correlated variables characterizes them as robust tools in environmental studies.

On the other hand, models based on ANNs have evolved considerably in recent years, incorporating more flexible architectures and optimization techniques such as genetic algorithms, wavelet transforms and hybrid strategies inspired by nature, improving performance against nonlinear and highly noisy datasets [107,159]. Its applicability has extended from the prediction of individual parameters to multivariable estimates of water quality, including dissolved oxygen and organic pollutants. Also, models based on fuzzy logic have gained relevance due to their ability to handle uncertainty (inherent in environmental data). Indeed, research in Saudi Arabia, India, and China has shown that fuzzy logic, when integrated with neural networks or evolutionary algorithms, exhibits suitable performances in regions with fragmented or scattered data [98,189,190]. Techniques such as Extreme Learning Machine (ELM), in combination with RF, have also been used to extend their applicability in environments with high hydrological variability [125].

Additionally, algorithms such as SVM, SVR, and RF remain highly popular due to their effectiveness in estimating water quality parameters. These methods offer non-parametric solutions that learn directly from observed data, facilitating uncertainty management and contextualized interpretation of pollution patterns [27].

The bibliometric analysis also reflects a sustained increase in integrating XAI techniques in research oriented to water quality prediction, to interpret interpreting the internal behavior of complex models and facilitating transparent decision-making. Techniques such as SHAP and LIME have been implemented to overcome the "black box" of ML/DL algorithms, allowing researchers to assess and visualize the influence of each predictor variable on the results [83,105,191]. XAI has been shown to improve model reliability and provide useful explanations for water quality management [106,192,193]. A prominent example is [194], in which an interpretable learning framework based on SHAP and RF is developed, applied to hydrodynamic scenarios. This approach allowed us to understand the impact of environmental variables on water quality, reinforcing the usefulness of XAI in complex studies of aquatic systems.

Figure 9 (a) summarizes the predictive models identified in this study, classifying and organizing the AI/ML/DL algorithms most commonly used in freshwater quality prediction studies, both for surface water and groundwater systems. The percentage of the models with the highest applicability in predicting water quality in both surface and groundwater is presented in Figure 9(b).

Finally, building reliable predictive models involves following a rigorous workflow that includes database consolidation, raw data preprocessing, proper predictive algorithm selection, model training, and validation. Recent literature underscores that each stage is critical to ensure the accuracy and robustness of the predictive model [48,195,196,197]. In particular, the selection of the algorithm must be aligned with the nature of the data and the specific objectives of a given study, ensuring interpretable, adaptable and relevant results for environmental decision-making.

RQ2. Which AI/ML/DL algorithm allows a better estimate of water quality?

The literature review identified a wide range of AI/ML/DL models developed both as standalone approaches and hybrid frameworks to enhance water quality prediction. Defining a single robust predictive model remains a challenge for researchers. However, certain algorithms stand out for their efficiency. For instance, the Light Gradient Boosting Machine (LightGBM) has emerged as a highly effective option due to its ability to process large datasets, fast training speed, and optimized architecture designed to minimize computational cost while maximizing predictive accuracy. Recent studies have reported accuracies exceeding 90% in the estimation of physicochemical parameters [198,199,200]. This technique is particularly notable for its consistent performance in water bodies with high variability.

The XGB algorithm has demonstrated remarkable performance in classifying both surface and groundwater quality, achieving accuracy levels close to 89% [201,202]. Its integration with XAI techniques such as SHAP enables the interpretation of key indicators such as zinc, nitrates, and chlorides, thereby improving the transparency of predictive models [201].

Time-series-based models, particularly LSTM architectures and their hybrid variants such as LSTM-CNN and CEEMDAN-LSTM, have achieved accuracies ranging from 90% to 93% by capturing complex dynamic patterns [6,44,86,114,153,203]. The incorporation of Transfer Learning further enhances performance by enabling efficient adaptation under data limitations [99,101,102]. Although MLP models have a simpler structure, they yield accuracies between 85% and 89% in sequential scenarios. Optimizing their parameters through genetic algorithms has improved precision in hydrologically realistic contexts [195,204,205,206].

Finally, hybrid approaches that combine multiple algorithms have reached accuracies above 92%, standing out for their adaptability to outlier values and their generalization capacity [155,207]. In summary, although no single model consistently outperforms across all contexts, algorithms that integrate interpretability, deep architecture, and hybrid strategies have delivered more accurate and reliable results. The selection of the most suitable model depends on the type of data, the analytical objectives, and the specific conditions of the water system under study.

RQ3. What limitations have been identified in the use of AI/ML/DL for water quality prediction?

A relevant aspect identified in this work is the study of spatial and temporal variations in the evaluation of water quality in the context of missing data, primarily due to missing data caused by measurement system failures, operational errors, environmental phenomena, and the non-continuous sampling frequency of water quality data. These limitations constrain the availability of reliable datasets for classification and evaluation purposes. In this context, Time Series Analysis (TSA) models combined with machine learning approaches, particularly Long Short-Term Memory (LSTM) networks have proven highly effective in addressing these challenges and estimating future values of water quality parameters (WQP) or water quality indices (WQI) based on historical data [208,209,210].

It was observed that DL models face various challenges compared to traditional physical models for water quality prediction. These include the complexity of internal structure and parameter adjustment, reliance on large data sets for effective training, and a lack of physical constraints, which can make it difficult to explain prediction results. Similarly, obtaining high-reliability data for certain water quality parameters can be difficult, limiting the applicability of deep learning approaches [211]. However, these challenges have begun to be overcome with the development of hybrid models and the use of interpretable approaches [26,86,112,212,213]. In general, the accuracy of the predictions of AI/ML/DL models is influenced by the availability and quality of historical data, and the models developed can be sensitive to variations in environmental conditions over time.

Recent literature has shown that some of these limitations can be mitigated through the use of LSTM models combined with Transfer Learning (TL) techniques [99,101], particularly instance-based approaches such as TrAdaBoost [94]. This model inherits the strengths of both LSTM and TL, offering powerful capabilities to capture long-term dependencies in time series and the flexibility to leverage related knowledge from complete datasets to fill large-scale consecutive data gaps [94]. Notably, prediction performance can be further enhanced by applying wavelet transforms to suppress noise in time-series signals, serving as an optimization mechanism for predictive modeling [87]. In this regard, model performance depends heavily on how variables are cleaned, transformed, and selected. For this reason, optimization algorithms such as Particle Swarm Optimization (PSO) [214,215], among others [112,116,216], are widely used.

Finally, variations in model performance caused by seasonal changes, extreme events, or point-source contamination which can significantly affect prediction accuracy have been addressed using Generative Adversarial Networks (GANs). These networks enable the simulation of abnormal or extreme conditions that are poorly represented in real datasets [161].

RQ4. What emerging variants currently exist in AI/ML/DL models for estimating water quality?

A current trend in water resource management using AI/ML/DL is the development of hybrid algorithms, which have gained relevance as a strategy to improve the accuracy of water quality parameter estimation. For instance, models based on variational mode decomposition optimized by the sparrow search algorithm (SSA-VMD), combined with Bidirectional Gated Recurrent Units (BiGRU), have achieved over 96% efficiency in the case of Qiandao Lake, China [217]. These hybrid approaches enable the capture of both nonlinear patterns and temporal dynamics in hydro-environmental data.

Besides, the Transformer architecture, has recently gained significant attention in deep learning research due to its superior performance over state-of-the-art LSTM models in time series forecasting and prediction tasks [27]. Finally, there is growing interest in enhancing the interpretability of machine learning models, given their “black box” nature. In this regard, Explainable AI (XAI) techniques have been implemented to identify the relative impact of water quality parameters on model predictions. Recent literature highlights successful applications of SHAP in models estimating salinity, dissolved oxygen, and the prediction of heavy metal concentrations, contributing to more informed decision-making in water management [162,163,218].

RQ5. What are the key water quality indicators used to assess natural water sources?

Water quality assessment in natural sources relies on a set of physicochemical and biological parameters that characterize both the environmental status and the suitability of water for various uses. According to the scientific literature reviewed, the most frequently used parameters in predictive studies employing AI/ML/DL techniques include dissolved oxygen (DO), biochemical oxygen demand (BOD), chemical oxygen demand (COD), total suspended solids (TSS), temperature, pH, electrical conductivity (EC), chlorophyll-a (Chl-a), nitrates, phosphates, and coliform bacteria. These indicators are considered conventional and are regulated by international environmental standards due to their relevance in characterizing both surface and groundwater bodies.

Beyond these essential parameters, there is a growing trend in the literature toward incorporating complementary variables such as heavy metals (e.g., arsenic, copper, lead), nutrients (NO₃⁻, NO₂⁻, NH₄⁺, PO₄³⁻), and trace metals (Fe, Mn, Zn, Cu, Cr), which enhance the discrimination between different water quality categories. Additionally, the integration of hydrogeological, meteorological, land use, and socioeconomic variables has proven valuable in enriching predictive models by capturing the influence of external factors on water body dynamics [127,219,220] In this regard, meteorological, land use, socioeconomic, and hydrogeological variables [221], for example, help illustrate how human activities can alter the export of chemical elements through changes in vegetation cover ultimately affecting water quality [222].

5. Contribution and Future Work

The review of the scientific literature performed in this work for the 2015-2024 period, through the integrated methodology of bibliometric and systematic review (B-SLR), delved into the AI/ML/DL techniques for the prediction of freshwater quality (surface water, groundwater). It also allowed us to identify key authors in the field, research gaps, emerging trends and models with high predictive performance. The main contribution of this study lies in its methodological approach that allowed identifying the most relevant works applying AI/ML/DL models in both surface waters and groundwater quality-related studies. It highlights an evolution of the models from traditional approaches such as ANN to hybrid and explainable architectures such as XGB, LSTM-CNN and SHAP.

The results of this research confirmed the findings of the study [58], which has received over 452 citations and reviewed predictive models for river water quality within the 2000–2020 timeframe. However, by applying the B-SLR methodology, the present study significantly expanded the scope of analysis—updating the literature to 2024, identifying emerging trends, and incorporating studies on freshwater groundwater as a potential source of drinking water. This broader and more integrative approach enabled a deeper thematic mapping of the field, revealing underexplored areas and offering a more comprehensive understanding of current scientific developments. Regarding groundwater quality, the study [107] of the time frame (1994-2022) highlighted that ANNs are the most used in water quality modeling and that nitrate is the most studied parameter. Complementing these insights, the present work identifies a transition from more traditional approaches such as ANNs to more sophisticated models such as XGB, LightGB, and hybrid models with LSTM and CNN. Finally, and as future research, we propose to expand the spectrum of water quality parameters considered in predictive models, incorporating emerging pollutants such as microplastics, pharmaceuticals and persistent organic compounds which are of current concern [193] in water management.

6. Conclusions

This research demonstrates that the Bibliometric-Systematic Literature Review (B-SLR) approach constitutes a robust methodology for analyzing the state of the art in highly dynamic scientific fields, such as water quality prediction using AI/ML/DL techniques. By integrating the structured rigor of systematic review with the analytical depth of bibliometric analysis, the B-SLR enabled the identification of domain-specific trends, thematic mapping of knowledge, assessment of scientific impact, and detection of gaps in the literature—offering a more comprehensive, precise, and context-aware understanding of the field.

The findings reveal that ensemble models (e.g., Bagging, Boosting), deep neural networks (LSTM, CNN, MLP), and hybrid approaches have overcome the limitations of conventional methods, delivering greater accuracy, adaptability, and the ability to handle incomplete or nonlinear data. The integration of explainable artificial intelligence (XAI) techniques, such as SHAP and LIME, has facilitated the development of more transparent and reliable models, enhanced result interpretation and supporting informed decision-making. Accordingly, the selection of the optimal predictive model depends on multiple factors, including the type of water body, geographic context, data availability, and the specific objectives of the study. In this regard, hybrid and interpretable models emerge as the most promising alternatives for addressing current challenges in water quality prediction.

Regarding the methodological approach employed in this study, the application of B-SLR enabled the refinement of an initial database of 1,822 articles into a final corpus of 274 highly relevant publications, through automated procedures and manual validation. This process ensured the quality and relevance of the analyzed studies, reinforcing the reliability of the results obtained. Furthermore, a detailed classification was achieved, covering the most frequently used algorithms, the types of water bodies studied, key quality indicators, and the methodological limitations faced by predictive models. Ultimately, this work establishes a replicable and scalable methodological foundation for future research.