Applying Artificial Intelligence (AI) Innovative Tools for Ecological Research and Monitoring of Transitional Water Ecosystems: A Systematic Review

Abstract

Transitional water ecosystems exhibit pronounced spatio-temporal variability and increasing anthropogenic pressures, posing substantial challenges for ecological monitoring and management. Artificial Intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL), has emerged as a powerful framework for addressing the structural complexity of these systems. This systematic review synthesizes peer-reviewed studies applying ML and DL to ecological research and monitoring in transitional waters. A structured search of the Scopus^® database was conducted up to 31 December 2024, and studies were screened according to predefined eligibility criteria and PRISMA 2020 guidance; methodological quality was appraised using a structured assessment framework. Ninety-six studies met the inclusion criteria. Regression was the most frequent analytical task (44.1%), followed by classification (36.2%) and clustering (19.7%), with water quality monitoring representing the dominant thematic domain. Tree-based and kernel-based ML models prevailed overall, whereas DL architectures increased markedly after 2020, particularly in remote sensing and high-dimensional applications. Despite methodological heterogeneity and variable validation practices, the evidence indicates that ML and DL approaches effectively accommodate non-linearity, data heterogeneity, and scale mismatches typical of transitional waters. Standardized validation strategies and improved model interpretability remain essential for robust ecological inference and operational implementation.

1. Introduction

1.1. Historical Development of AI

AI is commonly defined as the discipline concerned with designing computational systems that display goal-directed behaviour in complex environments [1]. The evolution of AI has opened new frontiers in aquatic ecosystem research. In recent years, ML and DL have revolutionized many scientific fields, including aquatic ecology [2].

Over the past decade, AI has advanced rapidly, driven by (i) algorithmic innovation, (ii) increased computational power, and (iii) the widespread availability of open-source neural-network software frameworks [3]. These developments have accelerated adoption across multiple domains, including medicine [4], biology [5], finance [6], manufacturing [7], engineering [8], and entertainment [9].

During the 1990s, ensemble approaches (e.g., boosting and bagging) and RF became influential methods for data analysis and prediction, complementing and, in some settings, outperforming classical statistical baselines. Contemporary ML is commonly framed in terms of supervised and unsupervised learning paradigms, as well as reinforcement learning for sequential decision-making, implemented through algorithms ranging from RF and neural networks to clustering techniques and modern deep-learning architectures [10].

Machine learning seeks to develop models that excel in making data-driven predictions, setting itself apart from traditional approaches that rely on specific assumptions about data generation processes [11]. In ecological applications, ML is typically implemented through a modelling workflow that can include (i) data transformation and feature engineering, (ii) supervised or unsupervised learning depending on label availability and study objectives, and (iii) optimization-based calibration (e.g., parameter estimation and hyperparameter tuning) under explicit validation designs. Representative examples include data transformation workflows [12,13], clustering-based ecological stratification [14], and optimization frameworks for parameter or hyperparameter selection [15,16].

Deep learning uses ANN models, composed of multiple interconnected layers, and represents a highly effective computational method [17]. ML underpins several operational AI technologies used in ecological monitoring, including automated classification and anomaly detection, and data-driven forecasting such as autonomous driving systems have significantly improved applications like image recognition, speech processing, and language translation [18]. Unlike traditional ML, which typically depends on human expertise to manually identify relevant data features—often challenging when dealing with complex or unclear features—deep neural networks automatically identify critical patterns within datasets. Furthermore, deep learning is increasingly utilized in ecology and evolutionary biology for diverse tasks, including citizen science initiatives, environmental monitoring, genetic sequencing, population genetics, and phylogenetic studies.

In the last decade, deep neural networks (DL) have seen renewed growth, primarily driven by advances in graphics processing unit (GPU) technology. These developments have significantly improved performance across various domains, including image analysis, language processing, and ecological modeling.

In transitional water ecosystems ecology and evolution, the adoption of ML and DL has spurred significant advances. AI has been applied to automated species identification from imaging and acoustic recordings [19,20,21,22,23], and to the extraction of organismal traits or behavioural proxies from high-frequency sensor, video, or bio-logging data [24,25,26,27,28]. Additionally, these approaches have been instrumental in predictive ecosystem modeling, contributing substantially to our understanding and management of these vital environments. For example, ML and DL have been used to fill missing links in ecological networks (e.g., Ref. [29]), integrate with traditional mechanistic models [30,31], approximate differential equations [32,33]. AI has also been used to build species distribution models by learning relationships between occurrence data and environmental covariates, including both correlative ML baselines and DL-based spatial predictors [34,35,36,37].

Despite their growing popularity, the complexity of ML and DL models often leads them to be perceived as “black boxes,” raising concerns about their explainability and relationship with traditional statistical models. However, recent work on XAI and emerging causal-inference approaches has begun to complement purely predictive modelling by improving transparency and supporting hypothesis generation. However, these methods do not automatically confer causal interpretability, and their ecological validity depends on study design, confounder control, and careful model diagnostics.

Integrating ML, DL, and traditional statistics has created a powerful methodological framework, transforming complex data analysis in AI. This interdisciplinary synergy continues to drive innovation in ML and data science.

1.2. Why a Review on AI Applied to Transitional Water Ecosystems?

Transitional water ecosystems are ecologically and socio-economically significant, supporting high biodiversity and providing essential ecosystem services such as storm protection, water filtration, and fisheries support. However, these ecosystems are increasingly vulnerable to climate change, anthropogenic pressures, and pollution, which threaten their biodiversity and ecological functions [38].

Integrating AI methods—particularly ML, DL, and remote sensing—has strengthened ecological research in transitional waters by enabling early detection of harmful algal blooms, habitat mapping, and quantitative assessment of environmental stressors [39]. More broadly, AI-based predictive analytics can support conservation and management by extracting actionable signals from large, heterogeneous environmental datasets and by anticipating climate- and human-driven impacts [40].

This review analyzes methodologies, challenges, and future opportunities for AI in transitional water ecosystems, contributing to scientific discourse on improving conservation and sustainable management in these environments.

Transitional water ecosystems pose a distinctive set of methodological challenges that limit the effectiveness of traditional statistical and mechanistic modelling approaches. These systems are characterized by strong spatio-temporal non-stationarity, abrupt regime shifts driven by episodic events (e.g., storms, freshwater inflows, heatwaves), and non-linear interactions among physical, chemical, and biological processes. Moreover, ecological responses are often modulated by scale mismatches between drivers (e.g., watershed-scale nutrient loading) and observations (e.g., point-based monitoring stations), as well as by highly heterogeneous and partially observed data streams.

From a data perspective, research in transitional waters must frequently contend with:

• high-dimensional predictor spaces arising from remote sensing, multi-sensor platforms, and omics-derived variables;
• missing, irregularly sampled, or noisy observations, particularly in long-term monitoring programs;
• limited labeled data, especially for biological states or rare events (e.g., hypoxia, harmful algal blooms);
• complex response surfaces that violate linearity, normality, and independence assumptions commonly required by classical statistical models.

These characteristics motivate the adoption of AI-based approaches not merely for their predictive accuracy, but because they offer algorithmic mechanisms capable of approximating complex, non-linear mappings, integrating heterogeneous data sources, and operating under weaker distributional assumptions than traditional parametric models. The relevance of AI in this context therefore lies in its alignment with the intrinsic structure of transitional-water data, rather than in generic claims of robustness or computational efficiency.

1.3. Origins of ML and DL in Relation to Statistics

ML and DL have their origins deeply rooted in the principles of statistics and AI. Since the 1950s (Figure 1), researchers have sought to develop computational models that learn from data, infer underlying patterns, and make autonomous decisions, thereby emulating aspects of human cognition.

Statistics has long provided the conceptual and methodological framework for data analysis and the modeling of complex phenomena. In addition to Bayesian statistics, key concepts such as maximum likelihood estimation (MLE) and null hypothesis significance testing (NHST) were established in the early twentieth century. These traditional parametric approaches involve formulating a model for data generation and computing the likelihood of observing specific outcomes based on assumed parameters. Pioneers such as Fisher, Neyman, and Pearson advanced these methods by introducing confidence intervals (CI) and p-values for parameter estimation and hypothesis testing, approaches that, however, recent critiques have highlighted recurring methodological risks—such as data leakage, inadequate treatment of spatio-temporal dependence, limited external validity, and insufficient ecological interpretability—particularly when models are deployed beyond the calibration domain [41,42,43,44].

Initially, practical statistical methods required simplistic models to ensure tractable probability calculations. However, the advent of early computers and novel numerical techniques like Markov Chain Monte Carlo (MCMC) methods [45] enabled the adoption of more complex parametric models. This progress paralleled advancements in ecological analysis [46]. Nonetheless, the inherent complexity of natural systems [47] often necessitated basic and inflexible statistical models, as computing likelihoods for intricate models remained mathematically challenging. Consequently, conclusions drawn from conventional statistical techniques frequently depended on oversimplified model assumptions [11].

The modern paradigm of ML emerged in part due to the availability of large datasets and the exponential increase in computational power. The 1990s witnessed a surge of innovation, with algorithms such as boosting, bagging, and RF surpassing traditional statistical models in flexibility and predictive performance. These developments firmly established ML as a fundamental approach to data analysis across diverse scientific disciplines and practical applications.

DL, which originated in the 1980s, initially faced limitations due to computational constraints and insufficient training data. Early neural network models showed promise but were restricted in scope. It was not until the advent of modern deep neural network architectures and the widespread use of GPUs that DL achieved its current potential. Throughout these developments, statistics has continued to play a crucial role by providing the theoretical underpinnings for model validation, result interpretation, and uncertainty quantification. Statistical concepts such as Bayesian inference, information theory, and sampling techniques have significantly influenced the evolution and application of ML algorithms.

Today, the integration of ML, DL, and statistical methodologies has culminated in a sophisticated methodological ecosystem, revolutionizing our approach to complex problems in data analysis and AI. This dynamic interplay among complementary disciplines continues to drive innovation and progress in ML and data science.

Evolution of AI in Transitional Water Ecosystems

Over the last two decades, the adoption of Artificial Intelligence methodologies in transitional water ecosystem research has followed a progressive and application-driven trajectory, closely linked to advances in data availability, monitoring technologies, and ecological research needs.

Early applications, emerging in the early 2000s, primarily relied on relatively simple machine-learning and neural-network models to support regression-based tasks, such as the estimation of water-quality variables (e.g., light attenuation coefficients, chlorophyll-a concentration) and the development of empirical ecological indicators. These studies represented the first attempts to move beyond purely mechanistic or linear statistical approaches by exploiting data-driven relationships in complex aquatic environments.

Between approximately 2010 and 2014, the diffusion of tree-based ensemble methods and kernel-based algorithms (e.g., Random Forests and Support Vector Machines) marked a substantial methodological shift. These approaches proved particularly well suited to transitional water ecosystems, where non-linear responses, strong interactions among predictors, and heterogeneous data structures are common. During this period, AI applications expanded to include species distribution modelling, aquaculture productivity assessment, and habitat suitability analysis in lagoons and estuarine systems.

From 2015 onwards, the integration of machine-learning models with remote-sensing data significantly broadened the spatial and temporal scope of ecological analyses. Ensemble methods and supervised classifiers were increasingly applied to land-use mapping, trophic status classification, and large-scale habitat assessments in coastal and transitional waters. This phase was characterized by the consolidation of AI as an operational tool for ecosystem monitoring rather than a purely experimental methodology.

More recently (from approximately 2020 to the present), advances in deep learning have enabled the analysis of high-dimensional and weakly structured data sources, including satellite imagery, underwater video, acoustic recordings, and high-frequency sensor time series. Convolutional and recurrent neural network architectures are now widely used for automated species detection, fine-scale habitat mapping, and short-term forecasting of ecological variables such as algal blooms and water-quality dynamics. These developments have facilitated near-real-time monitoring and the handling of data volumes that would be impractical using traditional analytical approaches.

In parallel, increasing attention has been devoted to model diagnostics, validation, and interpretability, particularly in response to concerns regarding the opacity of complex AI models. Recent studies have therefore explored the use of explainable AI techniques to support ecological interpretation, hypothesis generation, and decision-making, while acknowledging that explainability does not equate to mechanistic understanding.

Overall, the evolution of AI in transitional water ecosystem research reflects a shift from exploratory, regression-focused applications toward integrated, data-intensive frameworks capable of addressing the intrinsic spatio-temporal complexity of these systems. This progression underscores that the relevance of AI in this domain lies not in generic methodological novelty, but in its capacity to align with the specific structural and ecological challenges of transitional waters.

2. ML and DL in Transitional Water Ecosystem Research

Transitional water ecosystems (e.g., coastal lagoons, estuaries, deltas, and coastal wetlands) are characterized by pronounced spatio-temporal heterogeneity and strong coupling among physical, chemical, and biological processes. Salinity gradients, hydrodynamics, episodic disturbances, and multiple anthropogenic pressures (e.g., eutrophication, contaminant inputs, habitat alteration) generate complex, often non-linear system responses. In this context, AI—operationally discussed here through ML and DL—is increasingly adopted to (i) extract information from heterogeneous data streams (in situ sensors, remote sensing products, acoustic/video monitoring, environmental DNA (eDNA), and omics), (ii) support ecological forecasting and early-warning systems (e.g., hypoxia events, harmful algal blooms), and (iii) enable scalable monitoring and decision support. This section provides a methodological framing explicitly aligned with the aims of the present review, which compares AI approaches used in transitional waters according to task type, model family, application domain, and publication trends.

2.1. ML

ML is a pivotal technology underpinning many advances in the understanding and management of transitional water ecosystems. Unlike DL, which primarily leverages deep neural networks, ML encompasses a broader array of algorithms enabling computers to learn from data and generate predictions or decisions. In practice, ML approaches are commonly organized into supervised and unsupervised learning paradigms, as well as reinforcement learning for sequential decision-making, each offering distinct strategies for modelling complex environmental systems.

A defining characteristic of many ML methods is their algorithmic focus on optimizing predictive performance through the minimization of a broad loss function, which does not necessarily correspond to a probabilistic model of the observed data [11,48]. This paradigm reduces reliance on strict distributional assumptions (typical of classical inference) and has motivated non-parametric strategies for uncertainty assessment. In this context, bootstrap procedures have become widely used to estimate confidence intervals for parameters and predictions in both statistical and ML settings [49]. Similarly, cross-validation—partitioning data into training and test subsets to quantify generalization error—has become central to ML evaluation and is now feasible at scale due to modern computational resources [50,51].

In transitional waters, ML is widely applied to support core operational and scientific goals, including the prediction of algal blooms, monitoring of fish migration, assessment of water quality, and detection of pollution [52]. Within the analytical taxonomy adopted in this review, these applications are typically expressed as regression (e.g., estimating continuous indicators such as chlorophyll-a, turbidity, dissolved oxygen, nutrients), classification (e.g., bloom/non-bloom, ecological status classes, presence–absence outcomes), and, less frequently, clustering (e.g., grouping stations or periods by similarity to detect latent regimes).

ML can rapidly analyze large ecological datasets, identify trends, and deliver timely predictions compared to purely manual or low-dimensional approaches; however, its effectiveness depends on data quality, feature definition, and appropriate algorithm tuning, particularly under the strong space–time variability typical of transitional systems [53].

From a methodological perspective, the suitability of specific ML algorithms for transitional water ecosystems is closely linked to the structural properties of the underlying data rather than to generic performance claims. For instance, tree-based ensemble methods such as RF are particularly well adapted to ecological datasets characterized by non-linear responses, strong interaction effects, and multicollinearity among predictors—conditions that frequently arise in transitional waters due to coupled physical–biogeochemical processes.

The recursive partitioning mechanism of RF enables the approximation of complex response surfaces without requiring explicit specification of functional forms or interaction terms, making it effective for modeling threshold-driven ecological dynamics (e.g., hypoxia onset, eutrophication tipping points). Furthermore, the bootstrap aggregation and random feature selection steps reduce variance and mitigate overfitting in high-dimensional predictor spaces, which are common when integrating remote sensing, in situ monitoring, and ancillary environmental data.

Similarly, kernel-based methods such as Support Vector Machines (SVM) are well suited for transitional-water applications where sample sizes are limited but the feature space is high-dimensional, as often occurs in biological monitoring or water-quality assessment based on multi-sensor observations. By maximizing the margin between classes or fitting flexible regression functions in transformed feature spaces, SVMs can capture non-linear ecological relationships while maintaining controlled model complexity.

These algorithmic properties explain the widespread adoption of ML models in transitional-water research and clarify that their relevance arises from their ability to accommodate data heterogeneity, non-linearity, and scale mismatches intrinsic to these ecosystems.

2.2. DL

DL, a subset of ML, has transformed approaches to complex problems frequently encountered in transitional water ecosystems. DL relies on ANN models composed of multiple interconnected layers that learn hierarchical representations directly from data. This capacity is especially advantageous when relevant ecological information is embedded in high-dimensional or weakly structured inputs (e.g., satellite imagery, hyperspectral products, underwater video, passive acoustics, and high-frequency sensor time series), where manual feature engineering may be limiting.

In aquatic and transitional-water contexts, DL methods have been used to detect and classify marine species using imagery and sonar [54], and more broadly to support event forecasting (e.g., toxic algal blooms) and ecosystem health assessment in data-intensive workflows. These approaches can enhance predictive accuracy and support biodiversity-informed management by extracting non-linear relationships and complex patterns that are difficult to model with simpler formulations. At the same time, DL applications remain constrained by (i) the requirement for large, well-annotated training datasets, often difficult to obtain in transitional waters due to logistical constraints and natural variability; and (ii) challenges in model interpretability, which can limit scientific inference and operational uptake when decision-making requires transparent justification [55].

2.3. Advantages and Disadvantages of AI Applications

The application of AI (particularly ML and DL) to transitional water ecosystems offers substantial potential alongside notable challenges. On the one hand, AI can process large volumes of environmental data rapidly, enabling near-real-time monitoring and the timely synthesis of information from multiple sources [56]. This capability supports the detection of latent ecological relationships that may remain undetected under traditional analytical workflows [55], improves predictive performance for ecosystem dynamics when models are appropriately designed and validated [57], and strengthens conservation and management strategies by automating biodiversity monitoring, habitat assessment, and detection of environmental stressors [58].

On the other hand, the reliability of AI outputs is strongly dependent on the availability, quality, and representativeness of training data; biased or incomplete datasets can produce unreliable predictions and misleading inferences, particularly under the non-stationary conditions typical of transitional waters [57]. Furthermore, many high-performing models—especially DL architectures—exhibit limited transparency, complicating ecological interpretation and stakeholder trust [55]. Finally, the computational demands of model development and deployment can be substantial, with implications for accessibility and sustainability [56]. These methodological constraints, together with broader ethical and ecological risks associated with rapid deployment and misapplication, motivate a cautious and explicitly ecological framing of AI use in transitional-water monitoring and decision support [59].

2.4. Supervised and Unsupervised Learning

Artificial intelligence (AI) has become increasingly important in studying and managing transitional water ecosystems. A key methodological distinction in this domain concerns supervised versus unsupervised learning approaches (Figure 1). Supervised learning relies on labeled datasets to learn a mapping between predictors and outcomes, and it is widely used for tasks such as aquatic species classification, water quality forecasting, and organism distribution modelling based on historical observations [60,61,62]. In contrast, unsupervised learning does not require labeled data and aims to identify latent structure within multivariate datasets; in transitional waters, it is particularly useful for clustering sites or communities according to shared environmental characteristics and for exploratory analysis of time series to detect ecosystem changes [63,64]. In this review, these paradigms provide an organizing lens for synthesizing the methodological choices and application domains reported in the selected literature.

2.4.1. Supervised Learning

• What can be done

Building on the supervised-learning paradigm introduced in Section 2.4, labeled ecological observations can be leveraged to support prediction and decision-making in transitional water ecosystems. Typical applications include:

• Improve Classification Accuracy in Ecological Research: Supervised learning techniques such as Decision Trees, Random Forest, and SVMs have been successfully used to classify ecological datasets, improving prediction accuracy and efficiency [65].
• Optimize Model Selection and Hyperparameter Tuning: A comparative study of supervised learning algorithms has shown that tuning hyperparameters can significantly impact model performance, making it essential for improving prediction outcomes [66].
• Apply Supervised Learning to Predict Ecological Events: Machine learning models have been applied to predict ecological phenomena, such as mass mortality events in marine ecosystems, demonstrating the practical application of supervised learning in conservation research [67].
• Enhance Model Generalization and Adaptability: Studies suggest that incorporating multi-modal data in supervised learning models can improve generalization and adaptability in diverse ecological applications.
• Leverage Ensemble Learning for Improved Performance: Combining multiple supervised learning models using ensemble techniques such as bagging and boosting can enhance ecological modeling outcomes by reducing variance and bias [68].
• Use Supervised Learning for Real-time Monitoring: AI-driven models are increasingly being integrated with remote sensing and environmental monitoring systems to enable real-time ecological assessments [69].

• How can it be done

To implement supervised learning robustly in ecological research, a structured workflow is typically required, including algorithm selection, data preparation, model optimization, and rigorous evaluation. Key steps include:

• Choosing the Right Algorithm: Different supervised learning algorithms, such as Decision Trees, Random Forest, ANN models, and SVM models, can be applied to ecological datasets. Each method has unique advantages depending on the type and structure of the data [65].
• Data Preprocessing and Feature Selection: To improve accuracy, supervised models require proper data cleaning, normalization, and feature selection. Methods such as Principal Component Analysis (PCA) and Recursive Feature Elimination (RFE) help refine datasets before model training [70].
• Hyperparameter Tuning for Optimization: Model performance can be enhanced through hyperparameter tuning techniques like Grid Search, Random Search, or Bayesian Optimization, ensuring that models generalize well to new ecological data [70].
• Training with Labeled Data: Supervised learning requires a sufficient amount of labeled data for effective model training. Large-scale datasets, such as GPS telemetry data in ecological studies, improve classification accuracy [71].
• Evaluating Model Performance: Performance evaluation metrics, such as Accuracy, Precision, Recall, F1-score, and Area Under the Curve (AUC), help assess the reliability of supervised learning models in ecological research [70].
• Using Ensemble Learning for Robust Predictions: Combining multiple models using ensemble learning techniques, such as Bagging, Boosting (e.g., XGBoost), and Stacking, enhances prediction stability and reduces errors [70].
• Applying Supervised Learning to Real-world Ecological Problems: Supervised learning models are used to classify animal behavioral states from environmental variables, predict species distribution, and analyze biodiversity conservation strategies [70].

The following table (

Table 1. Supervised learning model.

Model	Overview	Applications
Linear Regression	Linear regression predicts a dependent variable based on its linear relationship with one or more independent variables, forming the basis of many statistical analyses.	It is commonly used to model and predict continuous outcomes, such as estimating population growth rates based on environmental factors or assessing the impact of habitat changes on species distributions.
Logistic Regression	Logistic regression addresses classification tasks with two possible outcomes by estimating the probability that data belong to a specific category.	Ecologists frequently use this approach to forecast species distributions under diverse environmental conditions and to predict occurrences of wildlife disease outbreaks.
Decision Trees	Decision trees partition data into branches at decision nodes based on feature values to predict outcomes.	They are effective in both classification and regression tasks, such as categorizing habitat types based on physical characteristics or predicting species richness across different ecological zones.
RF	RF are an ensemble method that aggregates predictions from multiple decision trees to improve predictive performance and mitigate overfitting.	They are used for complex classification and regression problems, including modeling biodiversity patterns across landscapes or assessing the impact of climate change on species distributions. RF is widely used for its ability to model non-linear relationships and complex interactions among predictors; however, it is not intrinsically interpretable, and ecological interpretation requires the use of post hoc, model-agnostic explainability techniques.
SVM models	SVM models are powerful classifiers that identify a hyperplane which maximizes the margin between classes.	They are applied to various classification problems, such as differentiating between invasive and native species based on morphological or genetic features, distinguishing among types of plankton, or assessing habitat suitability for specific species. Their ability to handle high-dimensional data and perform well with limited training samples makes them particularly valuable in ecological studies.
Neural Networks and DL	Neural networks consist of interconnected nodes (neurons) arranged in layers, allowing them to capture intricate, non-linear patterns. Deep learning involves deeper, multi-layer networks, enhancing their analytical power.	These techniques are widely utilized in ecological studies to process large datasets. They are particularly effective for tasks involving image analysis, like species identification from camera traps, and modeling complex ecological dynamics. Notable DL Architectures: CNN models: are particularly well-suited for analyzing data with a grid-like topology (e.g., images). They automatically extract relevant features, making them ideal for tasks such as species identification from underwater imagery or habitat mapping using satellite data. RNN models and LSTM models: RNN models are designed for sequential data and can maintain state information across time steps, making them useful for analyzing time series data. LSTMs, a specialized type of RNN, effectively capture long-term dependencies, making them particularly adept at forecasting ecological events such as algal blooms or monitoring water quality over time.
Gradient Boosting Machines (GBM)	GBMs build ensembles of predictors sequentially, with each new predictor correcting errors made by its predecessors, thus enhancing overall prediction accuracy.	They are effective in various ecological forecasting and modeling tasks, such as predicting changes in ecosystem services under different management scenarios or varying environmental conditions.

) summarizes the primary supervised algorithms used in these fields.

In summary, supervised learning algorithms allow ecologists to effectively analyze large ecological datasets, turning complex information into actionable insights. These tools are vital for accurate ecosystem monitoring, supporting proactive conservation and sustainable management strategies in aquatic environments.

2.4.2. Unsupervised Learning

• What can be done

As outlined in Section 2.4, unsupervised learning methods support exploratory analysis when labels are unavailable or incomplete, a frequent condition in transitional water monitoring. In aquatic ecosystem research, these techniques can uncover hidden patterns and relationships, helping to identify ecological clusters and detect temporal or spatial shifts. Key applications include:

• Clustering Ecological Data: Unsupervised learning techniques such as density-based clustering (DBSCAN) and k-means clustering have been successfully used to identify ecological patterns in marine environments [72].
• Detecting Anomalies in Ecosystems: Unsupervised learning is effective in anomaly detection, identifying unusual ecological events such as extreme climate conditions and pollution spikes without the need for labeled training data [73].
• Dimensionality Reduction for Complex Ecological Models: Techniques like t-SNE and PCA help in reducing high-dimensional ecological data, making it easier to visualize and interpret complex environmental interactions [72].
• Identifying Hidden Structures in Biodiversity: Unsupervised learning methods have been applied to analyze plankton community structures and nutrient flux patterns, leading to new insights in marine biodiversity research [72].
• Feature Extraction for Environmental Monitoring: Unsupervised neural networks have been used to extract meaningful features from satellite and sensor data, enhancing environmental monitoring efforts [74].
• Improving Conservation Strategies: By clustering species distribution data, unsupervised learning has been instrumental in helping conservationists identify priority areas for protection [75].

• How can it be done

Implementing unsupervised learning effectively typically involves careful preprocessing, algorithm selection, and validation tailored to exploratory objectives. Key steps include:

• Preprocessing and Feature Selection: Before applying unsupervised learning, datasets should be preprocessed to remove noise, handle missing values, and standardize data for better clustering and pattern detection [76].
• Choosing the Right Clustering Algorithm: Different clustering methods, such as k-means, hierarchical clustering, and density-based clustering (DBSCAN), are commonly used in ecological studies to classify species or environmental regions [72].
• Using Dimensionality Reduction Techniques: PCA and t-SNE are valuable tools for reducing high-dimensional ecological data, making it easier to visualize and analyze underlying patterns [77].
• Anomaly Detection for Ecological Monitoring: Unsupervised learning models are effective in detecting anomalies in environmental data, such as pollution events or habitat degradation, by identifying deviations from normal patterns [73].
• Leveraging DL for Unsupervised Feature Extraction: Deep learning models, such as autoencoders, can automatically learn features from large ecological datasets, enhancing species identification and habitat classification [77].
• Evaluating Model Performance: Unlike supervised learning, unsupervised models require different validation techniques, such as silhouette scores and clustering validity indices, to assess the quality of the detected patterns [78].

The following table (

Table 2. Evolution of Artificial Intelligence.

Model	Overview	Applications
K-Means Clustering	K-Means is one of the most straightforward and widely used unsupervised learning algorithms. It partitions data into a predetermined number (k) of distinct clusters based on feature similarity.	It is commonly applied to ecological niche identification, grouping species according to habitat preferences, and analyzing biodiversity patterns.
Hierarchical Clustering	Hierarchical clustering generates a tree-like structure (dendrogram) of clusters rather than a single partition, which is useful for revealing hierarchical relationships among species or habitats.	This technique is employed in phylogenetic analysis, vegetation classification, and studies of genetic diversity within and across species.
PCA	PCA reduces the dimensionality of environmental data while retaining most of the variation present in the dataset, thereby simplifying complex datasets for easier interpretation.	It is used to visualize genetic variation across populations, analyze environmental gradients that affect species distribution, and reduce the complexity of large ecological datasets.
Self-Organizing Maps (SOMs)	SOMs represent unsupervised neural networks that transform complex input data into simplified, lower-dimensional visualizations.	SOMs are valuable in ecological classification tasks, mapping spatial patterns of species, and tracking temporal variations in environmental conditions.
Independent Component Analysis (ICA)	ICA separates a multivariate signal into additive, statistically independent components, making it a valuable tool for blind signal separation.	It is used in the analysis of remote sensing data, disentangling mixed signals in acoustic ecology, and examining complex environmental datasets.
DBSCAN (Density-Based Spatial Clustering of Applications with Noise)	BSCAN groups data points that are closely packed together while identifying points in low-density regions as outliers.	It is ideal for detecting spatial patterns, such as clusters of biodiversity hotspots or analyzing the spatial distribution of species under varying environmental conditions.

) summarizes key unsupervised algorithms commonly employed in aquatic ecology and evolutionary biology research.

These unsupervised learning models serve a variety of purposes, from elucidating ecological relationships and community structures to identifying critical conservation areas. For instance, dimensionality reduction and clustering are frequently used to delineate ecological patterns and eco-provinces, highlight dominant environmental drivers of change, and support conservation prioritization by identifying ecological hotspots and areas at risk under climate variability and anthropogenic pressures [72,73,74,75]. Considerations related to the interpretability of complex models are addressed in Section 2.5.5 [79].

2.5. Evaluating and Improving AI Models

As AI becomes more prevalent in ecological research, strong evaluation frameworks are essential to maintain model reliability, accuracy, and transparency. AI models often face challenges in generalization, especially in ecology, where complex interactions and dynamic behaviors add layers of difficulty. Overcoming these obstacles requires interdisciplinary strategies that integrate AI techniques with ecological expertise [55].

2.5.1. Model Evaluation

Assessing AI models in ecological research is crucial to ensure their predictive accuracy. While standard metrics like accuracy, precision, recall, and F1-score are widely used, ecological data often demand more specialized evaluation methods. Techniques such as cross-validation, uncertainty analysis, and domain-specific error assessments help determine the robustness and reliability of AI models [80].

2.5.2. Error Diagnosis and Analysis

AI models can exhibit systematic biases or errors due to incomplete datasets, incorrect assumptions, or improper model architectures. Conducting error analysis through confusion matrices, residual analysis, and adversarial testing helps diagnose weaknesses in model predictions. Furthermore, integrating AI with physical-based ecological models has been shown to improve prediction accuracy by incorporating fundamental environmental principles [57].

2.5.3. Model Optimization and Tuning

Hyperparameter tuning is a crucial step in AI model improvement. Techniques such as grid search, Bayesian optimization, and evolutionary algorithms allow researchers to fine-tune model parameters for optimal performance. Additionally, incorporating domain knowledge into model design can lead to more ecologically meaningful results, reducing overfitting and enhancing generalization [55].

2.5.4. Feature Analysis

Optimizing AI models requires effective feature selection and engineering to improve parsimony and computational efficiency while preserving ecological relevance. Choosing ecologically meaningful variables—such as habitat distribution, environmental factors, and species interactions—ensures that models remain scientifically valid. Techniques like PCA and RFE help streamline datasets by removing redundant or irrelevant variables [81].

2.5.5. Interpretability and Explainability

A critical methodological issue in the application of AI to ecological research concerns the distinction between interpretability and explainability, two concepts that are often conflated in the literature but refer to fundamentally different properties of models.

Interpretability denotes the extent to which a model’s internal structure and parameters can be directly understood by a human observer. Interpretable models are typically characterized by a transparent mathematical formulation or a limited number of parameters with clear semantic meaning, such as linear regression models or shallow decision trees. In these cases, the relationship between predictors and outcomes can be directly inspected and reasoned about without auxiliary tools.

Explainability, by contrast, refers to the use of post hoc analytical techniques designed to approximate, summarize, or probe the behavior of complex and intrinsically opaque models. Most machine-learning models widely adopted in transitional water ecosystem research—including Random Forests, gradient-boosting methods, and deep neural networks—are not intrinsically interpretable due to their ensemble-based or multi-layered architectures. In such cases, model-agnostic explainability methods (e.g., SHAP, LIME, partial dependence analysis) can be employed to identify influential predictors, characterize local sensitivities, or generate simplified representations of model responses.

Importantly, explainability does not confer interpretability in a strict sense, nor does it imply causal understanding. XAI methods provide descriptive insights into how a trained model behaves under specific data distributions, but their outputs remain contingent on the model structure, the training data, and the assumptions embedded in the explainer itself. Consequently, explanations derived from XAI should be interpreted as diagnostic or hypothesis-generating tools rather than as mechanistic representations of ecological processes.

In the context of transitional water ecosystems, this distinction is particularly relevant because management and policy decisions often require transparent justification. While explainable AI can enhance trust and facilitate communication with stakeholders, robust ecological inference still depends on study design, data quality, and the integration of domain knowledge, rather than on explainability techniques alone.

Accordingly, throughout this review, models such as RF, gradient boosting methods, and DL architectures are discussed as non-interpretable models, whose outputs may nonetheless be explored using explainability tools for hypothesis generation and model validation purposes.

2.5.6. A/B Testing and Continuous Monitoring

Deploying AI models in ecological applications requires continuous evaluation and refinement. A/B testing—comparing different model versions on real-world ecological data—helps identify the most effective approaches. Additionally, long-term monitoring through feedback loops ensures that AI models remain adaptive to environmental changes and evolving datasets [57].

By implementing these evaluation and improvement strategies, AI models can become more reliable and transparent, ultimately contributing to more effective ecological research and conservation efforts. Future developments should focus on integrating AI with traditional ecological modeling techniques while maintaining explainability and ethical considerations.

3. Scope and Objective

This review synthesizes the international peer-reviewed literature on the application of ML and DL to transitional water ecosystems. Specifically, we analyze 96 peer-reviewed studies (including journal articles, review papers, and peer-reviewed conference proceedings indexed in Scopus) and systematically classify them according to AI task type, AI model family, application field, and year of publication. The objectives are to (i) provide an evidence-based overview of how AI is currently used for monitoring and ecological assessment in transitional waters, (ii) identify methodological and thematic trends across the literature, and (iii) highlight recurring limitations and research gaps that should be addressed to support robust ecological inference and operational implementation. The remainder of the manuscript is organized to first describe the review methodology and then present and discuss the main patterns emerging from the classified studies.

4. Materials and Methods

4.1. Review Design and Reporting Framework

This study was conceived as a systematic literature review aimed at synthesizing peer-reviewed evidence on the application of AI—with a specific focus on ML and DL—to ecological research and monitoring in transitional water ecosystems. To ensure transparency and reproducibility, the review process was structured in sequential phases (identification, screening, eligibility assessment, and inclusion) and documented using a PRISMA-style flow scheme (Figure 2). The review questions and the analytical framework (classification by AI task type, AI model type, application field, and year of publication) were defined a priori and applied consistently throughout study selection and data extraction.

4.2. Eligibility Criteria

Studies were considered eligible if they met all of the following criteria:

• Population/system: the study addressed transitional water ecosystems (e.g., coastal lagoons, estuaries, deltas, coastal wetlands, or clearly transitional aquatic environments);
• Intervention/method: the study applied AI approaches, with a specific focus on ML and/or DL, including hybrid or ensemble strategies, to ecological or environmental datasets;
• Outcomes/aims: the study targeted ecological research and/or monitoring objectives in transitional waters (e.g., assessment, forecasting, mapping, biodiversity evaluation, pollution detection, habitat characterization, or closely related management applications);
• Document type and source: peer-reviewed publications indexed in Scopus providing sufficient methodological detail for extraction and classification, including journal articles, review papers, and peer-reviewed conference proceedings;
• Language: English-language publications;
• Accessibility: full text available.

Exclusion criteria were:

• Studies not focused on transitional waters (e.g., exclusively open-ocean, purely freshwater, or terrestrial systems without a transitional component);
• Papers not implementing AI/ML/DL methods (e.g., purely descriptive studies, conceptual perspectives without implementation, or exclusively mechanistic modelling without an AI component);
• Excluded document types: non-peer-reviewed or non-research content (e.g., editorials, commentaries, meeting abstracts) and other records lacking sufficient methodological detail for robust extraction; additionally, records outside the eligible document types above (e.g., book notes, short communications without methods) were excluded during eligibility assessment;
• Records without accessible full text and/or insufficient information to classify AI task/model/application domain.

No a priori time restriction was imposed to capture the temporal evolution of AI methods in this field.

4.3. Information Source and Search Strategy

The bibliographic search was performed in the Scopus^® database (Elsevier, Amsterdam, The Netherlands), selected for its broad coverage of peer-reviewed journals in environmental sciences and interdisciplinary research. The search was conducted on 31 December 2024. Searches were performed using a structured query string combining AI-related terms with descriptors of transitional water ecosystems. The query was applied to the title, abstract, and keywords as follows:

(TITLE-ABS-KEY (machine AND learning AND coastal AND lagoons) OR TITLE-ABS-KEY (machine AND learning AND transitional AND water)).

To enhance replicability, we explicitly report the screening constraints applied after retrieval. The Scopus^® query was executed without time limits and without restricting document type at the search stage. During eligibility assessment, we retained English-language, peer-reviewed publications relevant to transitional water ecosystems that implemented AI/ML/DL approaches and provided sufficient methodological detail for extraction. Eligible records included journal articles, review papers, and peer-reviewed conference proceedings indexed in Scopus. Records such as editorials, commentaries, meeting abstracts, and other items lacking adequate methodological transparency were excluded and are summarized among the full-text exclusion categories in the PRISMA-style flow scheme (Figure 2).

To reduce the risk of missing relevant studies not captured by keyword combinations, we complemented the database query with targeted manual screening of reference lists of included papers (backward citation searching) and, where appropriate, identification of closely related articles suggested by the database interface (citation-based discovery). All retrieved records were exported and managed in a structured screening workflow.

4.4. Study Selection Process (PRISMA 2020)

Study identification and selection followed the PRISMA 2020 (Figure 2) reporting framework and its standard flow diagram for new systematic reviews based on database searches. The Scopus^® search (executed on 31 December 2024) retrieved 102 records. No duplicate records were detected during de-duplication (records removed before screening: n = 0). All records (n = 102) were screened on title and abstract; one record was excluded at this stage (records excluded: n = 1). The remaining 101 reports were sought for retrieval and all were successfully retrieved (reports not retrieved: n = 0). Full-text eligibility was assessed for 101 reports; five reports were excluded for predefined reasons: (i) not focused on transitional waters (n = 2); (ii) no AI/ML/DL implementation (n = 1); (iii) ineligible document type or insufficient methodological detail for extraction (n = 2). Accordingly, 96 studies were included in the qualitative synthesis and subsequent classification.

4.5. Data Extraction and Coding Scheme

A standardized extraction matrix was used to collect key attributes from each included study and ensure consistent classification across the evidence base. Extracted items included: bibliographic information (authors, year), ecosystem type and geographic context, data source (e.g., in situ monitoring, remote sensing, acoustic/video, laboratory measurements), AI task type (regression, classification, clustering, and/or detection/segmentation where applicable), AI model family (e.g., tree-based models, kernel methods, neural networks, CNN/RNN/LSTM architectures, hybrid/statistical models), application field (e.g., water quality monitoring, habitat mapping, biodiversity assessment, pollution detection), and evaluation/validation approaches (e.g., cross-validation, train/test split, performance metrics). The extracted variables were then used to populate the classification tables (Appendix A) and to support the synthesis presented in the Results.

4.6. Quality Assessment and Risk of Bias

To enhance methodological transparency and align with PRISMA 2020 recommendations, a structured quality appraisal of included studies was conducted. Given the methodological heterogeneity of AI-based ecological research, a domain-adapted quality assessment framework was developed.

Each study was evaluated according to five predefined criteria:

• clarity of study objectives and ecological context;
• transparency of dataset description and preprocessing;
• appropriateness of AI model selection relative to research question;
• robustness of validation strategy (e.g., cross-validation, external validation, spatio-temporal independence);
• reporting of performance metrics and uncertainty measures.

Each criterion was scored on a three-level scale (low, moderate, high quality). Two reviewers independently assessed methodological quality, and discrepancies were resolved through discussion.

The quality appraisal was used to identify recurrent methodological limitations but was not employed to exclude studies from the qualitative synthesis.

4.7. Synthesis Methods

The synthesis combined descriptive and exploratory analyses. First, studies were summarized through frequency counts and distributions across task types, model families, application fields, and publication years. Second, multivariate exploratory methods (including cluster analysis and non-metric multidimensional scaling) were used to examine temporal patterns and thematic similarities among studies, supporting the identification of emerging methodological trends and application domains (

Table A1. Systematic classification of 96 articles.

Paper #	Regression	Classification	Clustering	Articles	Topic	Model Type	Year
1	X			Neural Network-Based Light Attenuation Model for Monitoring Seagrass Health [105]	Marine Ecology and Environmental Monitoring: Monitoring seagrass health and assessing light attenuation in coastal and estuarine environments. This can be further classified under coastal ecology and water quality monitoring.	Machine Learning Models, Neural Networks	2004
				Problem Type: Regression: The study focuses on modeling the light attenuation coefficient using different water quality parameters to monitor seagrass health. This type of problem is a regression problem since it aims to predict a continuous value—the light attenuation coefficient. Approach: The study uses regression techniques to predict the vertical light attenuation coefficient, which is crucial for monitoring seagrass health. Various AI models, including neural networks, linear regression, model trees, and SVMs, are compared to identify the most effective approach. Topic: The study uses neural network-based models (MLP, GRNN) to estimate the light attenuation coefficient from 13 water quality parameters. These models are compared with traditional linear regression models, model trees, and SVMs. Models Used: The study uses neural network-based models (MLP, GRNN) to estimate the light attenuation coefficient from 13 water quality parameters. These models are compared with traditional linear regression models, model trees, and SVMs. Publication Year: The article was published in 2004.
2		X		Transcriptome Profiles: Diagnostic Signature of Dolphin Populations [106]	Marine Biology and Environmental Genomics: Understanding dolphin population structure and the environmental impact on health through transcriptomic profiling. This can be further categorized under species population monitoring and environmental stress response.	Machine Learning Models, Genomic Analysis Tools	2010
				Problem Type: The problem addressed is a classification problem, aiming to determine the geographic origin of dolphin populations using transcriptomic data. Approach: The study used cDNA microarrays to analyze gene expression, and ANN models to classify individuals and discern population differences. Topic: Population genetics and the impact of environmental factors on dolphin populations in different coastal areas. Models Used: ANN models and cDNA microarray analysis. Publication Year: 2010
3	X			Application of a Random Forest algorithm to predict spatial distribution of the potential yield of Ruditapes philippinarum in the Venice lagoon, Italy [107]	Marine Ecology and Aquaculture Management: The focus is on sustainable aquaculture of Manila clams, classified under marine ecology and aquaculture management.	Machine Learning Models	2011
				Problem Type: The study addresses a regression problem, predicting the yield of Manila clams. Approach: A Random Forest algorithm was employed to develop a model that predicts clam yield based on a set of environmental factors. The model was calibrated using a training dataset and validated with independent data. Topic: The main topic is marine ecology and aquaculture management, with a focus on the Venice Lagoon. Models Used: Random Forest. Publication Year: 2011
4	X		X	Field characterization and data integration to define the hydraulic heterogeneity of a shallow granular aquifer at a sub-watershed scale [108]	Groundwater and Hydrogeology: Specifically, aquifer characterization, which can also be categorized under water pollution (water quality) (due to the leachate plume).	Machine Learning Models	2014
				Problem Type: The study involves both regression and clustering. Regression is used for estimating hydraulic properties, while clustering is used to identify distinct hydrofacies units. Approach: A combination of cone penetrometer tests (CPT), soil moisture and resistivity measurements (SMR), and relevance vector machines (RVM) was used to characterize aquifer heterogeneity and predict hydraulic properties. Topic: Hydrogeology, specifically characterizing a shallow granular aquifer to better understand groundwater flow and contaminant transport. Models Used: Relevance Vector Machines (RVM). Year of Publication: 2014
5	X			A Statistical Algorithm for Estimating Chlorophyll Concentration in the New Caledonian Lagoon [109]	Marine Ecology and Water Quality: Specifically focusing on the health of lagoon ecosystems, including coral reefs and phytoplankton biomass estimation. This falls under water quality monitoring and marine ecosystem health.	Machine Learning Models Statistical Models	2016
				Problem Type: The study deals with a regression problem to estimate chlorophyll-a concentration. Approach: A statistical algorithm combining a log-linear model and SVM for low and high chlorophyll concentration scenarios. The models were trained on match-up data from 2002–2010. Topic: The study is related to marine ecology and aims to improve the estimation of chlorophyll concentration in optically complex waters. Models Used: SVM and log-linear models, compared to the NASA OC3 algorithm. Year of Publication: 2016
6	X			Developments in Earth observation for the assessment and monitoring of inland, transitional, coastal and shelf-sea waters [110]	Water Quality and Environmental Monitoring: The focus is on monitoring water quality using EO, which can be categorized under water pollution and environmental assessment.	Machine Learning Models Statistical Models	2016
				Problem Type: The paper deals with a regression problem, focused on estimating different water quality parameters using remote sensing. Approach: The approach combines satellite remote sensing, empirical algorithms, physics-based models, and machine learning techniques to assess water quality over large geographic areas such as the Danube-Black Sea system. Topic: The study is about water quality monitoring and environmental assessment of inland, coastal, and shelf-sea water systems. Models Used: Neural networks, empirical models, and other remote sensing algorithms. Year of Publication: 2016
7	X			Climate reconstructions based on postglacial macrofossil assemblages from four river systems in southwestern Alberta [111]	Palaeoclimatology and Environmental Monitoring: The focus is on palaeoclimatic reconstruction and ecological assessment of riverine systems in Alberta, which falls under climate change studies and ecological history.	Machine Learning Models	2017
				Problem Type: The study involves a regression problem that uses macrofossil data to estimate temperature and precipitation over different periods. Approach: The MAXENT model was used for climate niche modelling and predicting the climatic optima of indicator taxa. The study used macrofossil data from various river systems to derive temperature and precipitation trends since the end of the Pleistocene. Topic: Palaeoclimatic reconstruction using macrofossil analysis to understand ecological changes in southwestern Alberta. Models Used: MAXENT (Maximum Entropy Modelling). Year of Publication: 2017
8		X		Multi-temporal Land Use Mapping of Coastal Wetlands Area using Machine Learning in Google Earth Engine [112]	Land Use and Coastal Ecosystem Mapping: The study falls under coastal ecosystem monitoring and land use mapping, with an emphasis on coastal wetland management and understanding land use changes over time.	Machine Learning Models Tree-Based Models SVMs	2017
				Problem Type: Classification problem aimed at mapping land use in coastal wetland areas. Approach: The study utilizes Google Earth Engine (GEE) with ten different machine learning algorithms to perform multi-temporal land use mapping of the Segara Anakan Lagoon, using various satellite images from Landsat and ASTER GDEM. Topic: Land use mapping and coastal wetland analysis. Models Used: Fast Naive Bayes, CART (Classification and Regression Tree) RF models, GMO Max Entropy, Perceptron (Multi-Class Perceptron), Winnow, Voting SVM, Margin SVM, Pegasos (Primal Estimated Sub-Gradient Solver for SVM), IKPamir (Intersection Kernel Passive Aggressive Method for Information Retrieval, SVM) Year of Publication: 2017
9	X	X		Machine learning predictions of trophic status indicators and plankton T dynamic in coastal lagoons [113]	Water Quality and Marine Ecology: The focus is on monitoring and assessing eutrophication and plankton dynamics in a coastal lagoon, falling under water pollution and marine ecosystem health.	Machine Learning Models	2018
				Problem Type: The study involves both regression (for predicting trophic index values) and classification (to understand the dynamics of plankton and the controlling mechanisms in the lagoon). Approach: The Random Forest (RF) model was used to predict the trophic index (TRIX) of the lagoon using available environmental data. The model also classified plankton dynamics to determine whether they were driven by nutrient availability (bottom-up) or grazing pressures (top-down). Topic: The study focuses on eutrophication assessment and plankton dynamics in the Ghar El Melh Lagoon, a coastal Mediterranean lagoon. Models Used: Random Forest (RF). Year of Publication: 2018.
10		X		Supervised machine learning in predicting multiphase flow regimes in horizontal pipes [114]	Flow Dynamics and Petroleum Engineering: The focus is on predicting multiphase flow regimes in pipelines, which falls under flow assurance and petroleum engineering.	Machine Learning Models Tree-Based Models SVM Neural Network Models	2019
				Problem Type: Classification problem aimed at predicting the flow regime in multiphase flow scenarios. Approach: Several supervised machine learning algorithms were compared to predict the flow regimes. The best performing model was Random Forest, with high accuracy and efficient training time. Topic: The study focuses on multiphase flow prediction for the petroleum industry, specifically predicting the flow regimes in horizontal pipes containing air, water, and oil. Models Used: Random Forest, Decision Tree, Logistic Regression, SVM, Neural Network MLP. Year of Publication: 2019
11		X	X	Hyperspectral change detection in wetland and water-body areas [115]	Wetland Conservation and Water Quality: The focus is on environmental monitoring, specifically detecting changes in wetland and water body areas, falling under water pollution and wetland conservation.	Machine Learning Models Hybrid Models	2019
				Problem Type: The study involves both classification (classifying pixels as changed or unchanged) and change detection using hyperspectral imagery. Approach: The proposed method involves three main steps: image differencing, pseudo training data generation using an Expectation Maximization (EM) algorithm, and classification with a Random Forest classifier to produce a change map. Topic: Wetland and water body monitoring using hyperspectral change detection to observe changes over time in the Shadegan wetland. Models Used: Random Forest classifier, with pseudo training data generated using the Expectation Maximization (EM) algorithm. Year of Publication: 2019
12	X			Prediction of algal chlorophyll-a and water clarity in monsoon-region reservoir using machine learning approaches [116]	Water Quality and Eutrophication Management: This falls under water pollution monitoring and eutrophication management for reservoirs.	Machine Learning Models Statistical Models	2020
				Problem Type: The study involves a regression problem, predicting chlorophyll-a and water clarity (Secchi depth) in a reservoir. Approach: The researchers used MLR, SVM, and ANN to predict chlorophyll-a concentration and water clarity. The models were validated using cross-validation, with SVM showing the best performance in terms of predictive accuracy, based on metrics like RMSE, MAE, and R2. Topic: The study focuses on water quality and eutrophication management in a monsoon-region reservoir. Models Used: MLR, SVM, ANN. Year of Publication: 2020
13	X			Using Machine-Learning Algorithms for Eutrophication Modeling: Case Study of Mar Menor Lagoon (Spain) [117]	Water Pollution and Eutrophication: This study falls under water pollution monitoring and eutrophication management for coastal lagoons.	Machine Learning Models	2020
				Problem Type: The study involves a regression problem, predicting chlorophyll-a concentrations in the Mar Menor Lagoon. Approach: The researchers used MLNN and SVR models with various feature selection algorithms to predict Chl-a concentrations. The models were trained and validated with daily data collected from water quality campaigns. Topic: Eutrophication modeling in the Mar Menor Lagoon, focusing on improving the understanding of nutrient dynamics to manage eutrophication. Models Used: MLNN and SVR. Year of Publication: 2020
14	X		X	Optimum Design of a Seawater Intrusion Monitoring Scheme Based on the Image Quality Assessment Method [118]	Groundwater Monitoring and Coastal Management: The focus is on seawater intrusion in coastal aquifers, falling under water pollution and groundwater monitoring.	Machine Learning Models, Optimization Models: Genetic Algorithm, Hybrid Models	2020
				Problem Type: The study involves regression using KELM to estimate the groundwater model and optimization with a genetic algorithm to optimize the monitoring well locations. Approach: The methodology includes building a numerical simulation model for groundwater, using KELM as a surrogate to reduce computational load, and employing a genetic algorithm to optimize the monitoring scheme. Topic: The study focuses on seawater intrusion monitoring and optimization of groundwater monitoring networks. Models Used: KELM (Kernel Extreme Learning Machine) and Genetic Algorithm. Year of Publication: 2020
15	X	X		Pinna nobilis in the Mar Menor coastal lagoon: A story of colonization and uncertainty [119]	Biodiversity and Conservation: The study falls under species monitoring and conservation in the context of environmental change and habitat degradation. Specifically, it focuses on the effects of environmental factors on the population of a vulnerable species.	Machine Learning Models Tree-Based Models	2020
				Problem Type: The study involves classification to predict presence or absence and regression to estimate the probability of occurrence of Pinna nobilis in the Mar Menor lagoon. Approach: Random Forest models were employed to estimate the spatial distribution of P. nobilis across different time periods and identify the main environmental factors affecting its colonization. Topic: The study focuses on species distribution of the Mediterranean fan mussel and the environmental impact on its population in the Mar Menor lagoon. Models Used: Random Forest for classification and regression tasks. Year of Publication: 2020
16		X		SuperFish: A Mobile Application for Fish Species Recognition Using Image Processing Techniques and Deep Learning [120]	Biodiversity and Conservation: The focus is on fish species recognition to aid in marine conservation, fisheries management, and ecotourism. It falls under species identification and environmental monitoring.	Machine Learning Models Deep Learning Models Hybrid Models	2021
				Problem Type: The study deals with a classification problem to identify fish species from images. Approach: The approach involves a combination of traditional image processing steps (such as grayscale conversion, Gaussian blur, and contour extraction) followed by classification using various machine learning classifiers and deep learning (Inception-v3) models. The dataset included 1520 images representing 38 different fish species. Topic: The study focuses on the development of a mobile application for fish species recognition to support marine conservation, fisheries management, and ecotourism in Mauritius. Models Used: k-Nearest Neighbour (kNN), SVM models (SVMs), Neural Networks, Decision Trees, Random Forest, Inception-v3 deep learning model. Year of Publication: 2021
17		X		Using sentinel-2 satellite imagery to develop microphytobenthos-based water quality indices in estuaries [121]	Water Quality Monitoring and Ecosystem Assessment: The focus is on water pollution and ecosystem health, specifically using intertidal vegetation as indicators for water quality in estuaries.	Machine Learning Models	2021
				Problem Type: Classification problem involving the identification of intertidal vegetation types, including microphytobenthos and green macroalgae, using satellite imagery. Approach: The study used a Random Forest classifier applied to Sentinel-2 images to create a map of different vegetation types in French estuaries. NDVI was used as a proxy for vegetation biomass, and Sentinel-2’s spatial resolution allowed for detailed mapping. Topic: The study focuses on water quality assessment in estuaries, using microphytobenthos as a bio-indicator for environmental monitoring. Models Used: Random Forest classifier. Year of Publication: 2021
18		X		Mapping Spatial Distribution and Biomass of Intertidal Ulva Blooms Using Machine Learning and Earth Observation [122]	Water Quality and Marine Ecology: Focuses on macroalgal blooms and eutrophication in estuarine environments, falling under water pollution monitoring and coastal ecosystem health.	Machine Learning Models Hybrid Models	2021
				Problem Type: The study involves classification for mapping spatial distribution and regression for estimating biomass. Approach: The approach uses earth observation data (Sentinel-1/2 and Landsat) for mapping and NDVI for delineation of Ulva blooms. An ANN model is employed to predict biomass quantity using radar and optical reflectance data. Topic: Monitoring and management of macroalgal blooms in Irish estuaries to assess water quality and ecological status. Models Used: ANN. Year of Publication: 2021
19		X		Mapping and Quantification of the Dwarf Eelgrass Zostera noltei Using a Random Forest Algorithm on a SPOT 7 Satellite Image [123]	Biodiversity Monitoring and Conservation: The focus is on seagrass monitoring and environmental conservation in coastal areas, falling under habitat conservation and ecosystem health monitoring.	Machine Learning Models Tree-Based Models	2021
				Problem Type: Classification problem involving the use of Random Forest to classify habitats in the coastal lagoon. Approach: The approach involves using SPOT 7 satellite imagery and applying a Random Forest classifier to map the different types of habitats. Vegetation indices like NDVI were used to help in differentiating the habitats. Topic: The study focuses on seagrass monitoring and the mapping of dwarf eelgrass (Zostera noltei) in the Merja Zerga coastal lagoon to assess the impacts of environmental changes. Models Used: Random Forest Classifier. Year of Publication: 2021
20	X	X		Modeling fragmentation probability of land-use and land-cover using the bagging, random forest and random subspace in the Teesta River Basin, Bangladesh [124]	Land-Use and Environmental Management: The focus is on land-use and land-cover fragmentation and landscape ecology. It falls under environmental monitoring and land-use change analysis	Machine Learning Models Ensemble Models	2021
				Problem Type: The study involves classification for categorizing LULC types and regression for modeling fragmentation probability. Approach: The researchers used Bagging, RF, RSS, and their ensemble model to predict landscape fragmentation probability. They utilized remote sensing and GIS tools, with sensitivity analysis to understand the most influential parameters for modeling. Topic: The study focuses on landscape fragmentation and LULC change modeling in the Teesta River Basin, aiming to understand the environmental impacts of fragmentation and guide regional planning. Models Used: Bagging, Random Forest (RF), Random Subspace (RSS), and an Ensemble Model. Year of Publication: 2021
21		X		Applying Limnological Feature-Based Machine Learning Techniques to Chemical State Classification in Marine Transitional Systems [125]	Water Quality (Water Pollution and Quality Monitoring): The study addresses water pollution, specifically focusing on chemical state classification and eutrophication in marine ecosystems	Machine Learning Models Tree-Based Models	2021
				Problem Type: The study is a classification problem, aimed at classifying chemical states (eutrophication and contamination levels) in marine transitional systems. Approach: Machine learning models (LDA, CTree, NB, SVM) were applied to data related to nutrient, dissolved oxygen, chlorophyll-a, trace metals, and polycyclic aromatic hydrocarbons to produce chemical status classifications. The feature-based approach enabled the classification of the chemical state of estuarine and lagoon waters. Topic: The study is about water quality classification in marine transitional systems. Models Used: Linear Discriminant Analysis (LDA), Classification Tree (CTree), Naive Bayesian (NB), SVM Year of Publication: 2021
22	X	X		Integration of artificial intelligence–based LULC mapping and prediction for estimating ecosystem services for urban sustainability: past to future perspective [126]	Urban Expansion and Environmental Monitoring: The study falls under urban sustainability and land-use change analysis. It focuses on assessing the impacts of urban expansion on natural resources and ecosystem services	Machine Learning Models Hybrid Models	2021
				Problem Type: Classification for LULC mapping using SVM, and regression for predicting future LULC using ANN-CA. Approach: The study used SVM to classify LULC maps over different years and ANN-CA for future prediction. It also used RF and CART for sensitivity analysis. The future ESV was calculated based on the predicted LULC. Topic: The study focuses on urban sustainability and the dynamics of land-use change in the context of urban expansion in Saudi Arabia. Models Used: SVM, ANN-Cellular Automata (ANN-CA), Random Forest (RF), Classification and Regression Tree (CART). Year of Publication: 2021
23	X			A new approach to monitor water quality in the Menor sea (Spain) using satellite data and machine learning methods [127]	Water Quality and Marine Ecology: The study addresses water pollution by monitoring chlorophyll-a, an indicator of water quality and eutrophication in coastal areas	Machine Learning Models Deep Learning Models Hybrid Models	2021
				Problem Type: The study involves a regression problem for predicting chlorophyll-a concentration using satellite data. Approach: Sentinel-2 data is used along with machine learning models (RF, SVMRadial, ANN, and DNN) to estimate chl-a concentration in a cost-effective manner, providing near-real-time information. The models were evaluated under different feature selection scenarios, and RF achieved the best performance. Topic: The study focuses on water quality monitoring in the Menor Sea, specifically by estimating chlorophyll-a concentrations to track eutrophication. Models Used: Random Forest, SVMRadial, ANN, and DNN. Year of Publication: 2021
24		X		Classification of Boulders in Coastal Environments Using Random Forest Machine Learning on Topo-Bathymetric LiDAR Data [128]	Coastal Erosion and Habitat Monitoring: The focus is on coastal erosion prevention and the monitoring of marine habitats through the classification of boulders that contribute to natural coastal defenses and marine ecosystem health	Machine Learning Models Tree-Based Models	2021
				Problem Type: Classification problem involving the detection and classification of boulders in coastal environments using topo-bathymetric LiDAR data. Approach: A Random Forest classifier was used to analyze the LiDAR data. The study involved preprocessing the data to generate a point cloud of the seabed, followed by applying Random Forest to classify boulders based on different features extracted from the point cloud. Topic: The study focuses on boulder mapping in coastal areas to assess and manage natural coastal protection and habitat services. Models Used: Random Forest (RF) Year of Publication: 2021
25	X			Effect of changing weather conditions on Eastern Mediterranean coastal lagoon fishery [129]	Environmental Monitoring (Climate Change) and Fisheries: The study falls under climate change impact on fisheries and coastal ecosystems, focusing on the effects of changing weather conditions on fishery yields	Machine Learning Models Statistical Models	2021
				Problem Type: Regression problem involving the use of GLMs to model the effect of weather conditions on fishery catches. Approach: The study used Random Forest to select important predictors, followed by GLMs to predict fish catch. The models incorporated variables such as sea water temperature, maximum air temperature, wind speed, and cloudy days. Topic: The study focuses on the impact of climate change on coastal lagoon fisheries, analyzing how weather variations affect fishery yields. Models Used: Random Forest and Generalized Linear Models (GLMs). Year of Publication: 2021
26	X	X		Factors controlling iodine enrichment in a coastal plain aquifer in the North Jiangsu Yishusi Plain, China [130]	Water Quality (Water Pollution and Hydrogeochemistry): The study focuses on the enrichment of iodine in groundwater, which is related to water quality and pollution in coastal aquifers	Machine Learning Models Statistical Models	2021
				Problem Type: The study involves both classification (RF model) and regression (factor analysis) to understand and predict iodine concentrations in groundwater. Approach: The researchers used RF to identify important factors affecting iodine levels, and FA to understand the hydrogeochemical relationships. The study aims to identify the sources and geochemical processes of iodine in a coastal aquifer. Topic: The study focuses on iodine enrichment in groundwater in the North Jiangsu Yishusi Plain, China, and examines the hydrogeochemical processes controlling this phenomenon. Models Used: Random Forest (RF) and Factor Analysis (FA). Year of Publication: 2021
27		X		The application of game theory-based machine learning modelling to assess climate variability effects on the sensitivity of lagoon ecosystem parameters [131]	Environmental Monitoring (Climate Change Impact) and Ecosystem Sensitivity: The study addresses the impact of climate variability on lagoon ecosystems, specifically focusing on how changes in wind speed and temperature affect ecological sensitivity in coastal areas	Machine Learning Models Tree-Based Models Hybrid Models	2021
				Problem Type: The study involves a classification problem for modeling ecological parameters and sensitivity analysis using machine learning models. Approach: Multiple machine learning models, including RF, GBM, XGB, LR, SVC, and DT, were used to classify the ecological sensitivity of various parameters. The SHAP analysis was employed to assess the sensitivity of these parameters to year-to-year weather fluctuations, focusing on the influence of wind speed and air temperature on different biological and chemical components of the lagoon ecosystem. Topic: The study focuses on climate variability and its effects on the sensitivity of lagoon ecosystems, specifically targeting the Vistula Lagoon in the South Baltic. Models Used: RF, GBM, XGB, LR, SVC, DT, and SHAP. Year of Publication: 2021
28	X			Machine Learning Modeling Techniques for Forecasting the Trophic Level in a Restored South Mediterranean Lagoon Using Chlorophyll-a [132]	Water Quality (Water Pollution) and Marine Ecology (Eutrophication): The study focuses on water quality monitoring and eutrophication in coastal lagoons, specifically addressing the ecological status of a lagoon impacted by anthropogenic activities.	Machine Learning Models Hybrid Models	2021
				Problem Type: Regression was used to predict Chl-a concentrations in the North Lagoon of Tunis. Approach: The study used a NARX neural network to predict Chl-a concentrations, using key predictor variables identified through Random Forest feature selection. Data spanning from 1989 to 2018 were used for training, validation, and testing. Topic: The study addresses eutrophication in a coastal lagoon, focusing on the impact of urbanization and nutrient accumulation on water quality. Models Used: NARX Neural Network and Random Forest. Year of Publication: 2021
29		X	X	Zoning Seagrass Protection in Lap An Lagoon, Vietnam Using a Novel Integrated Framework for Sustainable Coastal Management [133]	Coastal Ecosystem Monitoring (Sustainable Coastal Management): The study addresses seagrass conservation in coastal lagoons, focusing on environmental protection and sustainable management practices	Machine Learning Models Decision-Making Models Hybrid Models	2021
				Problem Type: The study includes detection/segmentation (mapping seagrass using satellite imagery) and classification (ranking seagrass zones based on priority for conservation). Approach: The study integrated CatBoost for seagrass mapping and used MCE, fuzzy logic, and AHP to classify protection zones in a GIS framework. Topic: The research focuses on seagrass protection zoning in Lap An Lagoon, Vietnam, using an integrated framework to aid sustainable coastal management. Models Used: CatBoost, MCE, Fuzzy Logic, and AHP. Year of Publication: 2021
30	X		X	Satellite Multi/Hyper Spectral HR Sensors for Mapping the Posidonia oceanica in South Mediterranean Islands [134]	Coastal Ecosystem Monitoring: The study focuses on seagrass conservation and coastal habitat monitoring, particularly assessing the distribution and health of Posidonia oceanica in the context of anthropogenic impacts and climate change	Machine Learning Models Hybrid Models	2021
				Problem Type: The study involves detection/segmentation (to map the extent of seagrass) and regression (to estimate parameters like LAI). Approach: A combination of advanced machine learning algorithms was used to process satellite EO data for mapping and monitoring. The study integrated high-resolution (HR) multispectral and hyperspectral satellite sensors like Sentinel-2 MSI and PRISMA. Topic: Monitoring the Posidonia oceanica seagrass meadows in the South Mediterranean Islands to assess their health and extent under the impact of anthropogenic activities and climate change. Models Used: Random Forest, Gaussian Process Regressor, Kernel Ridge, Linear SVM, Partial Least Square, Support Vector Machine, and Maximum Likelihood. Year of Publication: 2021
31	X			On the runup parameterisation for reef-lined coasts [135]	Coastal Flooding and Resilience: The study focuses on predicting coastal flooding by modeling wave runup for reef-lined coasts, addressing the impact of changing sea levels and storm conditions on coastal vulnerability.	Machine Learning Models Hybrid Models	2022
				Problem Type: Regression was used to predict wave runup on reef-lined coasts. Approach: The study used a genetic programming approach to derive new parameterisations that consider wave conditions, reef geometry, and other hydrodynamic parameters. The SWASH model was used for simulating wave conditions and generating training data. Topic: The focus is on coastal flooding and the prediction of wave runup on reef-lined coasts, which is crucial for coastal resilience and mitigation planning. Models Used: Genetic Programming (GP) and SWASH numerical model. Year of Publication: 2022
32	X			Soft-ANN based correlation for air-water two-phase flow pressure drop estimation in a vertical mini-channel [136]	Fluid Dynamics and Pressure Drop Analysis: The study addresses pressure drop estimation in two-phase flow systems, which is important in various applications such as refrigeration, chemical engineering, and fluid transport systems	Machine Learning Models	2022
				Problem Type: The study is a regression problem aimed at predicting the pressure drop in air-water two-phase flow. Approach: The authors used an ANN model to estimate the frictional pressure drop in vertical mini-channels, incorporating dimensionless inputs like the Air-Reynolds number, Water-Reynolds number, and the ratio of air to water inertial forces. The ANN model was compared against existing correlations, showing superior performance. Topic: The focus is on estimating pressure drops in two-phase flow within vertical mini-channels to improve the accuracy of pressure drop predictions in fluid systems. Models Used: ANN. Year of Publication: 2022
33			X	Unsupervised Optical Classification of the Seabed Color in Shallow Oligotrophic Waters from Sentinel-2 Images: A Case Study in the Voh-Koné-Pouembout Lagoon (New Caledonia) [137]	Coastal Mapping and Habitat Classification: The study focuses on seabed mapping and classification, which falls under coastal habitat monitoring. The goal is to provide insights into seabed color variability and its ecological impact	Machine Learning Models Hybrid Models	2022
				Problem Type: The study involves clustering for seabed classification. Approach: The authors used Sentinel-2 satellite imagery along with the Lyzenga correction algorithm to preprocess the data, followed by k-means clustering for unsupervised classification of the seabed spectral data. Topic: The study aims to map and classify seabed colors in a tropical lagoon setting to improve understanding of the distribution of different seabed types and their impact on water quality. Models Used: k-means clustering and Lyzenga correction. Year of Publication: 2022
34	X			Towards hybrid modeling of the global hydrological cycle [138]	Hydrological Cycle and Water Resources: The focus is on global hydrological cycle modeling, which falls under the category of water resources management and hydrological modeling	Hybrid Models	2022
				Problem Type: The study addresses a regression problem focused on predicting different components of the global hydrological cycle. Approach: The study uses a hybrid modeling approach that combines neural networks (LSTM) with physical hydrological equations. The model, named H2M, simulates the global hydrological cycle by leveraging observational data to train the machine learning components, which replace some uncertain parameters in traditional models. Topic: The study aims to model the global hydrological cycle by predicting components like snow storage, soil moisture, and groundwater storage with a focus on maintaining physical consistency. Models Used: LSTM neural networks and traditional hydrological models. Year of Publication: 2022
35		X		Coastal Wetland Responses to Sea Level Rise: The Losers and Winners Based on Hydro-Geomorphological Settings [139]	Coastal Erosion and Habitat Shifts: The topic focuses on coastal wetland dynamics in response to sea level rise, which fits into the broader category of coastal habitat changes and sea level rise impacts	Tree-Based Models	2022
				Problem Type: The study is a classification problem where a RF model is used to classify different types of wetlands and predict their responses to sea level rise. Approach: The study uses machine learning (RF) to model wetland distribution based on the hydro-geomorphological settings of the study area, which is located in the Manning River estuary in New South Wales, Australia. Predictions were made under three sea level rise scenarios: low, moderate, and high. Topic: The study investigates the impact of sea level rise on the distribution of coastal wetlands, focusing on the winners (e.g., mangroves) and losers (e.g., freshwater swamps) of different wetland types under future scenarios. Models Used: Random Forest. Year of Publication: 2022
36	X	X		Assessing the Spatiotemporal Heterogeneity of Terrestrial Temperature as a Proxy to Microclimate and Its Relationship With Urban Hydro-Biophysical Parameters [140]	Urban Climate and Air Temperature: The study is related to urban climate modeling and the impact of urbanization on temperature, which fits into the broader category of air pollution and urban heat islands	Tree-Based Models	2022
				Problem Type: The study involves classification using Random Forest for LULC and regression to assess the relationship between LST and biophysical parameters. Approach: The research uses machine learning (Random Forest) for classification and advanced statistical techniques (Pearson’s correlation, Moran’s I) to explore relationships. Topic: The study focuses on the effects of urbanization on microclimate, specifically analyzing how different land types contribute to urban heat islands. Models Used: Random Forest. Year of Publication: 2022
37	X			Modelling Arctic coastal plain lake depths using machine learning and Google Earth Engine [141]	Climate Change and Water Resources: The study focuses on Arctic hydrology, particularly the effects of climate change on lake depths, which falls under climate change impact on freshwater resources	Machine Learning Models Tree-Based Models	2022
				Problem Type: The study is a regression problem focused on predicting lake depths using remote sensing data. Approach: Satellite imagery from Landsat-8 was used along with machine learning algorithms (RF, CART, and SVM) on Google Earth Engine (GEE) to predict bathymetric data. Random Forest was found to be the best-performing model. Topic: The focus is on lake depth modeling in the Arctic coastal plain, a region significantly affected by climate-induced changes, to better understand the effects of warming on water bodies. Models Used: Random Forest, CART, and SVMs. Year of Publication: 2022
38			X	Using a clustering algorithm to identify patterns of valve-gaping behaviour in mussels reared under different environmental conditions [142]	Coastal Habitat Monitoring and Behavioral Ecology: The study is related to coastal ecosystem monitoring, focusing on behavioral ecology of bivalves, specifically how mussels adapt their valve-gaping to varying environmental conditions in coastal ecosystems	Machine Learning Models	2022
				Problem Type: Clustering was used to analyze valve-gaping behavior in mussels. Approach: The k-means clustering algorithm was applied to time series data of mussel valve-gaping. Three rounds of clustering were conducted to explore daily variations, relationships with environmental factors like temperature, oxygen saturation, and chlorophyll levels, and intra-daily oscillatory patterns. Topic: The research aims to understand the physiological and behavioral adaptations of mussels, particularly valve-gaping behavior, in response to changing environmental conditions in transitional coastal ecosystems. Models Used: k-means clustering. Year of Publication: 2022
39			X	Smart Cameras for Coastal Monitoring [143]	Coastal Monitoring and Management: The topic falls under coastal erosion monitoring, shoreline management, and beach usage analysis to aid in sustainable coastal management.	Machine Learning Models	2022
				Problem Type: The study involves detection/segmentation for shoreline monitoring and beach usage tracking. Approach: A smart camera system integrated with machine learning and image processing techniques was deployed to quantify beach usage and track shoreline changes. The system is intended for long-term, continuous data collection that can aid in operational coastal management and provide insights into the morphological impacts of coastal processes. Topic: The study focuses on coastal monitoring, specifically for shoreline change detection and beach usage analysis, aimed at informing sustainable coastal management strategies. Models Used: The study utilizes machine learning algorithms combined with image processing, though the specific models are not detailed. Year of Publication: 2022
40	X			Near-infrared spectroscopy for prediction of potentially toxic elements in soil and sediments from a semiarid and coastal humid tropical transitional river basin [144]	Water Quality (Water Pollution): The study deals with contamination in river basins, specifically focusing on monitoring toxic elements in soils and sediments that may affect water quality	Machine Learning Models Chemometric Models	2022
				Problem Type: Regression for predicting PTE concentrations. Approach: Near-infrared (NIR) spectroscopy combined with chemometric models, including Random Forest and Partial Least Squares (PLS), to predict the concentrations of elements such as Al, Ti, Fe, La, and others. Topic: Monitoring potentially toxic elements in soils and sediments from the Ipojuca river basin, which is exposed to pollutants from agricultural and industrial activities. Models Used: Random Forest and Partial Least Squares regression. Year of Publication: 2022
41	X			Recurrent neural networks for water quality assessment in complex coastal lagoon environments: A case study on the Venice Lagoon [145]	Water Quality (Water Pollution): Specifically focused on eutrophication, which is one of the most common forms of water pollution affecting coastal environments	Neural Network-Based Models Tree-Based Models	2022
				Problem Type: Regression Approach: The study employs RNN variants (RNN, LSTM, GRU), CNN hybrids, and RF models to predict eutrophication based on water quality parameters. Topic: Eutrophication prediction in the Venice Lagoon, focusing on assessing water quality deterioration. Models Used: RNN models (RNN, LSTM, GRU), CNN-RNN hybrids, Random Forest. Year of Publication: 2022
42	X		X	Monitoring System of the Mar Menor Coastal Lagoon (Spain) and Its Watershed Basin Using the Integration of Massive Heterogeneous Data [146]	Water Quality (Water Pollution): The study deals with monitoring nutrient runoff, oxygen levels, and the ecological state of the lagoon, which is directly related to water quality and pollution	Machine Learning Models	2022
				Problem Type: Detection and monitoring of the environmental state. Approach: The study integrates a large volume of heterogeneous data (from sensors, satellite imagery, and hydrological networks) using a distributed and scalable system. The data are analyzed using classical and advanced data processing methods, including Machine Learning, for predictive and monitoring purposes. Topic: Environmental monitoring of the Mar Menor coastal lagoon. Models Used: Machine Learning techniques (specific models not mentioned). Year of Publication: 2022
43	X		X	Quantitative evaluation of the roles of ocean chemistry and climate on ooid size across the Phanerozoic: Global versus local controls [147]	Marine Sedimentology: Focuses on carbonate grain types (ooids) in shallow marine environments and the influence of environmental conditions on their growth.	Machine Learning Models Stochastic Models	2022
				Problem Type: Detection/Segmentation using machine learning and Monte Carlo simulations. Approach: The study utilized a CNN-based segmentation method to measure ooid sizes in petrographic images and used Monte Carlo simulations to explore the impact of different physicochemical parameters on ooid size. Topic: The study evaluates the roles of ocean chemistry, climate, and local environmental conditions on ooid growth across the Phanerozoic. Models Used: Fully Convolutional Networks (FCN) for segmentation and Monte Carlo simulations for parameter analysis. Year of Publication: 2022
44	X	X	X	Invited Review: Ecosystem services provided by grasslands in the Southeast United States [148]	Sustainable Agriculture and Ecosystem Services: The study addresses the contribution of grasslands to ecosystem services, which falls under the category of sustainable land management and ecosystem services	Machine Learning Models	2022
				Problem Type: Detection and monitoring of ecosystem services. Approach: Use of remote sensing, machine learning, and artificial intelligence to assess ecosystem services. Topic: Ecosystem services provided by grasslands in the southeastern United States, with a focus on their importance for sustainable agroecosystems. Models Used: Remote sensing, machine learning, and artificial intelligence (specific models not mentioned). Year of Publication: 2022
45			X	Scaling in size, time and risk—The problem of huge extrapolations and remedy by asymptotic matching [149]	Structural Engineering: Specifically related to concrete structures, fracture mechanics, creep and shrinkage, and structural safety.	Physics-based Theoretical Models	2023
				Problem Type: Scaling extrapolation related to structural size, time, and failure probabilities. Approach: Uses asymptotic matching and other theoretical methods rather than AI or machine learning, emphasizing physics-based models. Topic: Scaling of structural strength, creep and shrinkage of concrete, and the challenges of making predictions over large time and size scales. Models Used: Primarily theoretical models, including asymptotic analysis and fracture mechanics. Machine learning and deep learning are mentioned but noted as inadequate for this type of problem. Year of Publication: 2023
46	X			Estimation of Chl-a in highly anthropized environments using machine learning and remote sensing [150]	Hydrological and Water Resource Management: Specifically related to streamflow prediction, hydroelectric power management, and flood mitigation in the Amazon and Savanna regions.	Tree-based Models Distance-based Models Neural Networks Kernel-based Models	2023
				Problem Type: Regression (streamflow prediction). Approach: Used machine learning models to predict streamflow using rainfall data estimated by remote sensing and gauge stations. Topic: Streamflow prediction in the Brazilian Savanna and Amazon biomes to aid reservoir operations, hydroelectric power production, and flood mitigation. Models Used: Random Forest (RF), ANN, SVM, K-Nearest Neighbor (kNN). Year of Publication: 2023
47		X	X	Rirsk analysis of coastal areas: AN AI-based perspective using SAR DATA [151]	Coastal Flooding and Risk Management: The study falls under coastal erosion and flood risk analysis. It specifically addresses coastal vulnerability due to subsidence and extreme flooding events	Tree-Based Models	2023
				Problem Type: The study includes both detection/segmentation (for detecting subsidence and flooded areas) and binary classification (to classify pixels as “changed” or “unchanged”). Approach: The approach involves using SAR data from Sentinel-1 to conduct a Change Detection analysis and evaluate subsidence and flooding events in the Venice Lagoon. The Small Baseline Subset (SBAS) technique was used for interferometric analysis, and Random Forest was applied for the change detection classification. Topic: The study focuses on risk analysis of coastal areas by evaluating the effects of subsidence and flooding in the Venice Lagoon using SAR data and machine learning. Models Used: Random Forest (RF), used for change detection. Year of Publication: 2023
48		X	X	Machine and deep learning methods for Detection and Mapping of coastal wetlands of Crozon Peninsula (Brittany, France) used metric and sub-metric spatial resolution [152]	Coastal Wetlands Monitoring and Conservation: The study is related to coastal wetlands mapping, focusing on the use of remote sensing and machine learning techniques to detect and classify different wetland types.	Tree-Based Models Deep Learning Models	2023
				Problem Type: The study involves detection/segmentation and classification. It uses RF and CNN models to map wetland areas and classify various land cover types. Approach: The research used Sentinel-1 and Sentinel-2 images, combined with a Random Forest classifier and CNN to perform pixel-based and object-based classifications. The best results were achieved with the pixel-based Random Forest classification (kappa = 0.89, overall accuracy = 0.90, F1-score = 0.90). Topic: Mapping and classification of coastal wetlands for better understanding and preservation of these ecosystems. Models Used: Random Forest (RF) and CNN. Year of Publication: 2023
49			X	Machine Learning for Detection of Macroalgal Blooms in the Mar Menor Coastal Lagoon Using Sentinel-2 [153]	Water Quality (Water Pollution): The study is related to water pollution caused by eutrophication, which results in macroalgal blooms due to excess nutrient input	Machine Learning Models Explainability Tool	2023
				Problem Type: Detection and clustering. Approach: The approach combines satellite remote sensing using Sentinel-2 imagery, followed by machine learning techniques, including K-Means clustering and CART classification. Topic: Detection of macroalgal blooms in the Mar Menor coastal lagoon. Models Used: K-Means Clustering, CART for classification, and SHAP values for model interpretability. Year of Publication: 2023 Models Used: The study employs several machine learning models: Clustering Algorithms: K-Means and Cascade Simple K-Means. Classification Algorithms: SVM and Classification and Regression Tree (CART). The CART model provided the best results. Explainability: SHAPley Additive explanations (SHAP) values were used for algorithm explainability. Year of Publication: 2023
50	X	X		Ensemble habitat suitability modeling for predicting optimal sites for eelgrass (Zostera marina) in the tidal lagoon ecosystem: Implications for restoration and conservation [154]	Coastal Restoration and Conservation: Focuses on the restoration and conservation of eelgrass in coastal lagoon environments.	Machine Learning Models Tree-based Models Statistical Models Profile-based Models	2023
				Problem Type: The study addresses both classification and regression tasks for habitat suitability assessment. Approach: Ensemble habitat suitability model (EHM) was developed by combining results from ten models to predict the optimal sites for eelgrass conservation and restoration. Topic: Habitat suitability modeling for eelgrass (Zostera marina) in the tidal lagoon ecosystem. Models Used: Profile, regression, classification, and machine learning models, with machine learning models like RF, ANN, and GBM providing the best results. Year of Publication: 2023
51	X			A sophisticated model for rating water quality [155]	Water Quality (Water Pollution) and Coastal Management: The topic fits into categories like water pollution, coastal water quality assessment, and environmental monitoring for coastal and transitional waterbodies	Machine Learning Models Tree-based Models	2023
				Problem Type: Regression problem for water quality prediction. Approach: The study develops the Irish Water Quality Index (IEWQI) to assess water quality using indicators like salinity, dissolved oxygen, and nutrient levels. It employs RF for feature selection and MLR for sensitivity analysis. Topic: Coastal and transitional water quality assessment. Models Used: Random Forest for indicator selection and Multilinear Regression for sensitivity analysis. Year of Publication: 2023
52		X		Vegetation Dynamic in a Large Floodplain Wetland: The Effects of Hydroclimatic Regime [156]	Floodplain Wetland Dynamics: This study falls under the macro category of wetland management and conservation with a focus on vegetation dynamics. It could be categorized alongside topics such as wetland habitat transition and riverine ecosystem conservation	Machine Learning Models	2023
				Problem Type: Classification. Approach: The study used Landsat imagery and machine learning algorithms (RF) to classify four wetland habitats and employed GAMM to understand their dynamics over 34 years. Topic: Vegetation dynamics in a large floodplain wetland (East Dongting Lake) under the influence of hydroclimatic regime changes. Models Used: Random Forest for habitat classification and GAMM for investigating habitat dynamics. Year of Publication: 2023
53		X	X	A REVIEW AND TEST OF SHORELINE EXTRACTION TECHNIQUES [157]	Coastal Erosion and Management: The study addresses the monitoring and management of coastal areas through shoreline extraction, making it relevant to coastal erosion and land–sea boundary mapping	Machine Learning Models Statistical Models Unsupervised Learning Models	2023
				Problem Type: The study addresses both classification (using supervised and unsupervised methods) and detection/segmentation (for shoreline extraction). Approach: The study utilizes satellite images (Sentinel-1 and Sentinel-2) to extract shorelines using various classification techniques, including Random Forest, Maximum Likelihood, Minimum Distance, and K-means clustering. The accuracy of each method was validated against manually extracted shorelines. Topic: The study focuses on shoreline extraction from satellite remote sensing data for coastal monitoring and management. Models Used: Random Forest, Maximum Likelihood, Minimum Distance, K-means. Year of Publication: 2023
54		X		Coastal landscape classification using convolutional neural network and remote sensing data in Vietnam [158]	Coastal Landscape Management: The topic falls under coastal landscape classification and sustainable coastal management	Deep Learning Models	2023
				Problem Type: Classification of coastal landscapes. Approach: The authors used CNN models (CvNet) with remote sensing data from ALOS, NOAA, and multi-temporal Landsat satellite images to classify various coastal landscapes in Vietnam. The study included data preparation, CNN model training, and application for coastal landscape recognition. Topic: Coastal landscape classification in Vietnam. Models Used: CNN (CvNet) with different optimizers. Year of Publication: 2023
55		X		Machine learning classification of Austin Chalk chemofacies from high-resolution x-ray fluorescence core characterization [159]	Reservoir Characterization and Geochemistry: The study falls into reservoir characterization with a focus on understanding geochemistry, depositional environments, and reservoir quality in unconventional reservoirs	Deep Learning Models	2023
				Problem Type: Classification problem for chemofacies. Approach: The approach involves developing a semi-supervised chemofacies clustering using XRF core characterization, and then using a deep neural network model to predict chemofacies across multiple cores. Topic: Chemofacies classification in the Austin Chalk Group to better understand depositional environments and reservoir quality. Models Used: Deep Neural Network. Year of Publication: 2023
56		X		Research progress and prospects of hyperspectral remote sensing for global wetland from 2010 to 2022 [160]	Wetland Monitoring: The study addresses topics such as wetland vegetation classification, mangrove monitoring, salt marsh vegetation, and global wetland changes under coastal and aquatic ecosystems.	Machine Learning Models Deep Learning Models	2023
				Problem Type: Classification, focusing on wetland vegetation, mangroves, and salt marshes. Approach: Utilizes hyperspectral remote sensing with machine learning models for information extraction, feature extraction, and classification. Topic: Monitoring and mapping wetland ecosystems, focusing on vegetation and soil properties. Models Used: RF models, Decision Trees, SVMs, DL models. Year of Publication: 2023
57		X		Prediction of environmental factors responsible for chlorophyll a induced hypereutrophy using explainable machine learning [161]	Water Quality (Water Pollution), particularly related to nutrient enrichment and hypereutrophic conditions	Tree-based Models	2023
				Problem Type: Classification, aimed at predicting the transition between eutrophic and hypereutrophic states. Approach: The study uses the XGBoost model combined with SHAP to explain which environmental factors influence the water body’s state. The model includes environmental parameters such as nutrient levels, phytoplankton concentrations, and physical factors. Topic: Understanding and predicting the factors responsible for hypereutrophy in the Vistula Lagoon. Models Used: XGBoost for classification, combined with SHAP for interpretability. Year of Publication: 2023
58	X			Deep Learning-Based Time Series Forecasting Models Evaluation for the Forecast of Chlorophyll a and Dissolved Oxygen in the Mar Menor [162]	Water Quality (Water Pollution): Specifically focusing on hypoxia and eutrophication events in the Mar Menor coastal lagoon.	Machine Learning Models Deep Learning Models Hybrid Models	2023
				Problem Type: Regression Approach: The study used deep learning models to predict the weekly levels of chlorophyll a and dissolved oxygen in the Mar Menor, a coastal lagoon. The prediction was done using environmental monitoring data obtained from sensors and agro-climatic stations. Topic: Coastal lagoon monitoring, specifically targeting water quality parameters to assess hypoxia and eutrophication. Models Used: Time2Vec (BiLSTM and Transformer), CNN-BiLSTM, MDN-BiLSTM, TCN-BiLSTM, and Seq2Seq, among others. Year of Publication: 2023
59	X			Spatiotemporal characteristics of ozone and the formation sensitivity over the Fenwei Plain [163]	Air pollution; it focuses on surface ozone levels and pollution control strategies.	Tree-based Models	2023
				Problem Type: Regression Approach: The study uses a deep forest machine learning model to estimate ozone concentrations across the Fenwei Plain using high-resolution satellite and meteorological data. Topic: Investigation of ozone pollution in the Fenwei Plain, its drivers, and formation regimes using deep learning and satellite data. Models Used: Deep forest model (DF21). Year of Publication: 2023.
60		X		A comprehensive review of water quality indices for lotic and lentic ecosystems [164]	Water Quality (Water Pollution): The study falls under the category of assessing the quality of water resources and the impact of pollution on freshwater ecosystems.	Machine Learning Models Tree-Based Models Hybrid Model	2023
				Problem Type: Classification Approach: Using various water quality indices to assess the water quality of lotic and lentic ecosystems, while incorporating machine learning models to improve accuracy and reduce uncertainty. Topic: Water quality assessment in freshwater ecosystems (lotic and lentic). Models Used: XGB, RF, SVM, KNN, ANN, GNB. Year of Publication: 2023.
61	X			A hyperspectral inversion framework for estimating absorbing inherent optical properties and biogeochemical parameters in inland and coastal waters [165]	Water Quality (Water Pollution): The study focuses on monitoring inland and coastal water quality, including parameters like chlorophyll-a, total suspended solids, and dissolved organic matter	Hybrid Model	2023
				Problem Type: Regression, involving the estimation of multiple biogeochemical parameters and inherent optical properties. Approach: The study uses Mixture Density Networks (MDNs) to estimate parameters from hyperspectral satellite imagery. Topic: Remote estimation of optical properties and biogeochemical parameters in inland and coastal waters. Models Used: Mixture Density Networks. Year of Publication: 2023
62			X	Modeling the spatiotemporal heterogeneity of land surface temperature and its relationship with land use land cover using geo-statistical techniques and machine learning algorithms [166]	Urban Heat Island: Studying urban heat island effects in urban areas. Urban Expansion Climate Change: Impact of urban growth on land surface temperature.	Machine Learning Models Geospatial Models Hybrid Models	2022
				Problem Type: Detection/segmentation of LST clusters based on urban biophysical factors. Approach: The study employed geospatial, statistical, and machine learning methods, including SVM, LISA, MWA, and PCP. Topic: Impact of urban expansion on land surface temperature in Abha-Khamis Mushyet, Saudi Arabia, from 1990 to 2020. Models Used: SVM, Local Indicator of Spatial Associations (LISA), Mono-Window Algorithm (MWA), and Parallel Coordinate Plot (PCP). Year of Publication: 2023.
63	X			Improving ecological indicators of arid zone deserts through simulation [167]	Water Resources and Ecosystem Monitoring: The study focuses on monitoring ecological indicators in arid zones, including soil moisture, latent heat, and gross primary productivity, which aligns with broader topics like water resources management and ecosystem monitoring.	Hybrid Models Land Surface Models (LSMs)	2023
				Problem Type: Regression and simulation of ecological indicators. Approach: Combination of deep learning for data reconstruction, ecological simulation, and data assimilation. Topic: Improving ecological indicators in arid zones (Yarkant River Basin). Models Used: LSTM, Noah-MP, HRLDAS, GLDAS, EnKF. Year of Publication: 2023
64		X	X	Predicting priority management areas for land use/cover change in the transboundary Okavango basin based on machine learning [168]	Land Use Change and Planning: Focuses on land use/cover change, management areas, and planning in a transboundary region. Water Resources and Basin Management: Part of the study relates to ensuring sustainable water resources in the Okavango basin, which includes land management for maintaining water availability.	Machine Learning Models Hybrid Model	2023
				Problem Type: The study uses a combination of classification and detection approaches to model LULC change and determine PMAs in the Okavango basin. Approach: An ensemble of machine learning models is applied, using social-ecological data and LULC change transitions derived from satellite images to predict susceptible areas. Topic: Predicting areas susceptible to land use/cover change in the Okavango basin to guide transboundary land use planning. Models Used: Random Forest, Gradient Boosting Models, MaxEnt, Classification Tree Analysis, ANN. Year of Publication: 2023
65	X			Spatio-Temporal Dynamics of Total Suspended Sediments in the Belize Coastal Lagoon [169]	Water Quality (Water Pollution): Monitoring suspended sediment levels, their sources, and their impact on coastal lagoon ecosystems. Coastal Management: Addressing the effects of anthropogenic activities on coastal water quality and contributing to sustainable coastal management strategies.	Machine Learning Models Deep Learning Model Time-Series Analysis Model	2023
				Problem Type: Regression and time-series analysis for TSS estimation and anomaly detection. Approach: The study used machine learning models (RF, XGB, and DNN) to estimate TSS concentrations from Sentinel-2 satellite data and employed Bayesian methods for identifying temporal anomalies. Topic: Understanding the spatio-temporal dynamics of TSS in the Belize Coastal Lagoon. Models Used: RF, XGB, DNN, and BCD. Year of Publication: 2023
66	X			Evaluating the use of machine learning algorithms in environmental sensing for energy saving [170]	Coastal Management: The study focuses on the efficient management of sensors to protect and understand coastal ecosystems. Water Quality (Water Pollution): Monitoring variables like chlorophyll and turbidity is directly related to assessing water quality and pollution levels	Tree-Based Models Linear Models Hybrid Models Deep Learning Model	2023
				Problem Type: Regression, focusing on predicting future values of chlorophyll and turbidity. Approach: Several machine learning models were evaluated to predict the chlorophyll and turbidity values sensed by smart buoys in the Mar Menor lagoon. The study aims to reduce energy usage by predicting the sensed values, thus allowing the buoy to operate intermittently. Topic: Environmental sensing and energy saving in coastal lagoons using machine learning. Models Used: LR, Lasso, Elastic Net, KNN, SVM, MLP, CART, AdaBoosting, Gradient Boosting, Random Forest, Extra Tree. Year of Publication: 2023
67	X	X		Prediction model of type and band gap for photocatalytic g-GaN-based van der Waals heterojunction of density functional theory and machine techniques [171]	Photocatalysis: Focused on hydrogen production from water splitting	Machine Learning Models Tree-based Models	2023
				Problem Type: Classification and regression models were used for predicting heterojunction types and band gap values. Approach: Machine learning was applied to classify and predict characteristics of 2D vdW heterostructures. The models utilized descriptors based on chemical composition, structural features, and computational calculations. Topic: Photocatalysis for hydrogen generation through water splitting using 2D g-GaN-based vdW heterojunctions. Models Used: SVM, Adaboost, Random Forest, and XGBoost. Year of Publication: 2023
68	X	X		Venice lagoon chlorophyll-a evaluation under climate change conditions: A hybrid water quality machine learning and biogeochemical-based framework [172]	Water Quality (Water Pollution) with a specific focus on the effects of climate change on eutrophication and water quality in coastal lagoon environments.	Hybrid Model Machine Learning Models Tree-Based Models	2023
				Problem Type: Classification (Chl-a values into classes) and regression (predicting future Chl-a variations). Approach: Hybrid modeling using Random Forest and Multi-Layer Perceptron for classification and SHYFEM-BFM biogeochemical projections for future scenarios. Topic: Evaluating the impact of climate change on chlorophyll-a concentrations in the Venice Lagoon. Models Used: RF, MLP, and SHYFEM-BFM. Year of Publication: 2023
69		X	X	A hybrid modeling approach for detecting seasonal variations in inland Green-Blue Ecosystems [173]	Ecosystem Monitoring: Specifically, the study deals with monitoring inland Green-Blue Ecosystems, which includes aspects related to aquatic and terrestrial ecosystems. Water Resources: Given the focus on the influence of water and groundwater presence, it can also be categorized under water management and climate impact on ecosystems.	Hybrid Model Machine Learning Models Deep Learning Model	2024
				Problem Type: Classification and detection. Approach: Hybrid modelling approach combining deep learning (CNN with U-Net and ResNet-34) and traditional machine learning models (SVM and DT). Topic: Detecting and mapping seasonal variations in inland Green-Blue Ecosystems (GBE) in a Mediterranean climate zone using satellite data. Models Used: CNN, SVM, and Decision Trees. Year of Publication: 2024 (published online in December 2023).
70	X			Predictive modeling of nitrogen and phosphorus concentrations in rivers using a machine learning framework: A case study in an urban-rural transitional area in Wenzhou China [174]	Water Quality (Water Pollution), specifically focusing on nutrient pollution (e.g., nitrogen and phosphorus) in aquatic systems caused by urban and agricultural runoff	Machine Learning Models Ensemble Models Neural Network-Based Models	2024
				Problem Type: Regression. Approach: Use of Random Forest regression to model the spatiotemporal distribution of nitrogen and phosphorus in the river system, comparing its performance to other machine learning methods. Topic: Water quality monitoring, specifically focused on nitrogen and phosphorus pollution in an urban-rural transitional watershed. Models Used: Random Forest, Decision Tree Regression, AdaBoost Regression, K-Nearest Neighbor, ANN, Support Vector Regression. Year of Publication: 2024
71	X			Data-driven modelling for assessing trophic status in marine ecosystems using machine learning approaches [175]	Water Quality (Water Pollution) as it deals with assessing nutrient levels in marine environments and their effects, specifically eutrophication	Machine Learning Models Tree-based Models Regression Models	2024
				Problem Type: Regression (focused on predicting chlorophyll a concentration) Approach: Development of the Assessment Trophic Status Index (ATSI) model incorporating machine learning methods to enhance the accuracy of eutrophication assessment in marine ecosystems. Topic: Assessing trophic status in coastal and transitional waters (eutrophication). Models Used: XGBoost, DNN, RF, DT, ExT, KNN, SVM, GNB, MLR. Year of Publication: 2024
72	X			Toward a Digital Twin: combining sensing, machine learning, and data visualization for the effective management of a coastal lagoon environment [176]	Water Quality (Water Pollution): The study focuses on the prediction of chlorophyll and turbidity, indicators of water quality, and aims to monitor and mitigate eutrophication in coastal environments. Coastal Management: The Digital Twin framework for Mar Menor represents a novel approach to managing and preserving coastal ecosystems.	Machine Learning Models Tree-Based Models Hybrid Models	2024
				Problem Type: The study tackles a regression problem focusing on predicting chlorophyll and turbidity values in the Mar Menor lagoon. Approach: The approach includes integrating IoT sensors with machine learning models to create a Digital Twin for monitoring, predicting, and visualizing the lagoon’s environmental status. Topic: Coastal lagoon management and monitoring with a focus on eutrophication and water quality. Models Used: Several machine learning models were used, including tree-based models like Random Forest and Gradient Boosting, as well as other algorithms like SVM, KNN, and MLP. Year of Publication: 2024
73		X		Mapping surface sediment characteristics in enclosed shallow-marine environments using spatially balanced designs and the random forest algorithm [177]	Coastal erosion and sediment dynamics, as it focuses on mapping sediment characteristics and understanding factors influencing sediment deposition and movement, which are related to coastal changes.	Machine Learning Models	2024
				Problem Type: Classification and Regression Approach: Machine Learning-based sediment classification and prediction using Random Forest and spatial sampling designs. Topic: Mapping and predicting seafloor sediment characteristics in Port Phillip Bay, Australia. Models Used: Random Forest algorithm Year of Publication: 2024
74			X	How to disentangle sea-level rise and a number of other processes influencing coastal foods? [178]	Coastal Flooding: The study’s main focus is understanding and predicting coastal flooding events and their contributing factors	Machine Learning Models	2024
				Problem Type: Detection/Segmentation of various contributing phenomena to coastal flooding. Approach: Application of a decomposition method to sea-level data combined with machine learning-based hindcasting and operational modeling. Topic: Disentangling the factors leading to coastal flooding in Venice, specifically addressing why the flood in November 2019 was underestimated. Models Used: Machine learning for hindcasting, operational modeling. Year of Publication: 2024
75	X			Streamflow prediction based on machine learning models and rainfall estimated by remote sensing in the Brazilian Savanna and Amazon biomes transition [179]	Water Resources Management: Streamflow prediction. Flood Management: Specifically aimed at mitigating floods. Hydropower Production: Managing reservoir operations for hydroelectric power plants.	Machine Learning Models Tree-Based Models Hybrid Models	2024
				Problem Type: Regression, focusing on predicting streamflow for managing hydrological systems and reservoir operations. Approach: Various machine learning models were used to predict streamflow, including kNN, SVM, RF, and ANN. The models took time-lagged streamflow and average rainfall (from rain gauges and remote sensing products) as inputs. The Random Forest model showed the best performance in terms of accuracy. Topic: Streamflow prediction in the Brazilian Savanna and Amazon biomes transition, primarily for flood mitigation and efficient reservoir operation of hydroelectric plants. Models Used: kNN, SVM, RF, and ANN. Year of Publication: 2024
76			X	Estimating four-decadal variations of seagrass distribution using satellite data and deep learning methods in a marine lagoon [180]	Coastal Ecosystem Monitoring: The study falls under coastal ecosystem monitoring, specifically focusing on seagrass distribution in coastal lagoons.	Deep Learning Models Machine Learning Models	2024
				Problem Type: Detection/Segmentation Approach: Utilizes deep learning models (UNet and SegNet) for detecting seagrass from Landsat satellite imagery. Topic: Monitoring the distribution and health of seagrass in Swan Lake over four decades. Models Used: UNet and SegNet, with SegNet providing the best results. Year of Publication: 2024
77			X	Unsupervised clustering of catalogue-driven features for characterizing temporal evolution of labquake stress [181]	Seismic research category, specifically focusing on earthquake forecasting and stress evolution in fault zones.	Machine Learning Models Tree-Based Models Graph-Based Models	2023
				Problem Type: Unsupervised clustering. Approach: Extraction of catalogue-driven features from labquake experiments and clustering them using K-means to characterize different stages of stress evolution. Topic: Temporal evolution of labquake stress. Models Used: K-means, GMM, hierarchical clustering, spectral clustering. Year of Publication: 2024
78		X	X	Artificial Neural Networks for Mapping Coastal Lagoon of Chilika Lake, India, Using Earth Observation Data [182]	Environmental monitoring and biodiversity conservation, specifically focusing on coastal ecosystem monitoring of Chilika Lake, which can also be classified under coastal habitat mapping and climate change impact assessment	Tree-Based Model Support Vector Model Neural Network Model	2024
				Problem Type: Classification of land cover types. Approach: Multispectral satellite images from Landsat 8–9 were used, processed using machine learning techniques (ANN, Random Forest, SVM) to classify and analyze environmental changes in Chilika Lake. Topic: Environmental monitoring of the Chilika Lake coastal lagoon, focusing on biodiversity, habitat changes, and climate impacts. Models Used: ANN (MLP), Random Forest, SVM. Year of Publication: 2024
79	X			Towards better predicting the settling velocity of film-shaped microplastics based on experiment and simulation data [183]	Water Quality (Water Pollution), specifically focusing on microplastic pollution in aquatic environments and modeling their settling behavior.	Tree-Based Model Support Vector Model Automated Machine Learning Hybrid Models	2024
				Problem Type: Regression problem for predicting settling velocity. Approach: Empirical formula and machine learning models for predicting the settling velocity of microplastics using both experimental and simulation data. Topic: Settling velocity prediction of film-shaped microplastics in aquatic environments. Models Used: SVR, TPOT, XGBoost, Random Forest, Ensemble Learning. Year of Publication: 2024
80		X		Identification of the Structure of Liquid–Gas Flow in a Horizontal Pipeline Using the Gamma-Ray Absorption and a Convolutional Neural Network [184]	Industrial process control and fluid mechanics, focusing on the monitoring and classification of multiphase flow regimes	Deep Learning Models	2024
				Problem Type: Classification of flow regimes. Approach: Use of gamma-ray absorption to gather data on liquid–gas flow and applying a CNN (VGG-16) for classification. Topic: Identification of liquid–gas flow structures in pipelines. Models Used: VGG-16 CNN. Year of Publication: 2024
81	X			Reservoir horizontal principle stress prediction using intelligent fusion model based on physical model constraints: a case study of Daji Block, Eastern Ordos Basin, North China [185]	Reservoir geomechanics and oil and gas exploration, specifically focusing on stress prediction in shale formations and reservoir management	Tree-Based Models Support Vector Model Linear Regression Models Hybrid Models	2024
				Problem Type: Regression to predict horizontal principal stress and rock mechanical parameters. Approach: Use of laboratory test data combined with multiple machine learning models to create an intelligent fusion model for predicting stress and mechanical parameters. Topic: Prediction of horizontal principal stress in transitional shale reservoirs in the Ordos Basin, China. Models Used: KNN, SVM, DT, RF, XGB, MLR, and Fusion Model. Year of Publication: 2024
82		X	X	Flow Regime Classification Using Orientation-Independent Layered Spectral Clustering [186]	Nuclear thermal-hydraulics and reactor safety and focuses on fluid dynamics, specifically flow regime identification in two-phase flows	Clustering Models Layered Hybrid Approach	2023
				Problem Type: Classification and clustering. Approach: The study uses spectral clustering with a layered structure to classify flow regimes in inclined two-phase flows. This unsupervised approach helps avoid the subjectivity found in traditional classification techniques. Topic: Objective identification of inclined flow regimes in two-phase flows, focusing on overcoming subjectivity and inclination-based variability in traditional classification methods. Models Used: Layered Spectral Clustering. Year of Publication: 2023
83	X	X		A Multi-Factor Analysis Linking Environmental Stressors with Presence of Per- and Polyfluoroalkyl Substances (PFAS) in a Coastal Lagoon [187]	Water Quality (Water Pollution) and environmental monitoring, focusing specifically on the presence and behavior of Per- and Polyfluoroalkyl Substances (PFAS) in aquatic environments.	Tree-Based Models	2024
				Problem Type: Classification and Regression. Approach: Use of the CART algorithm to predict the occurrence of PFAS in the Indian River Lagoon based on environmental stressors. Topic: Investigating PFAS occurrence in a coastal lagoon and identifying key environmental stressors associated with PFAS presence. Models Used: Classification and Regression Trees (CART). Year of Publication: 2024
84		X		Seagrasses on the move: Tracing the multi-decadal species distribution trends in lagoon meadows using Landsat imagery [188]	Coastal ecosystem monitoring and environmental change assessment, focusing specifically on biodiversity and seagrass dynamics in a coastal lagoon.	Tree-Based Models	2024
				Problem Type: Classification (community and species levels). Approach: Seasonal multispectral satellite images (Landsat 5 and Landsat 8) combined with Random Forest to classify seagrass meadows and monitor changes over time. Topic: Multi-decadal monitoring of seagrass distribution in the Grado and Marano lagoon (Northeast Italy). Models Used: Random Forest (RF). Year of Publication: 2024
85	X			Estimating hourly air temperature in an Amazon-Cerrado transitional forest in Brazil using Machine Learning regression models [189]	Climate modeling and environmental monitoring, focusing specifically on temperature prediction in a tropical forest ecosystem	Tree-Based Models Support Vector Model Neural Network Model	2024
				Problem Type: Regression. Approach: The study used several machine learning models to predict hourly air temperature in the Amazon-Cerrado region, comparing their performance across both dry and rainy seasons. Topic: Predicting air temperature in an Amazon-Cerrado transitional forest. Models Used: Random Forest, Gradient Boosting Regressor, Support Vector Regressor, Multilayer Perceptron. Year of Publication: 2024
86		X		Using shell shape analysis based on landmarks to trace the geographical origin of the common cockle (Cerastoderma edule) [190]	Traceability and food safety, focusing on seafood origin determination and environmental monitoring related to cockle harvesting in coastal systems	Tree-Based Models Support Vector Model Neural Network Model Statistical Approaches	2024
				Problem Type: Classification. Approach: Landmark-based geometric morphometric analysis of cockle shells, using various machine learning models for classification. Topic: Determining the geographical origin of common cockle (Cerastoderma edule) from coastal systems in Portugal. Models Used: LDA, CVA, PCA, bgPCA, PLSD, CRT, LR, RF, GB, KNN, SVM, NNET. Year of Publication: 2024
87	X	X		Joint mining of fluid knowledge and multi-sensor data for gas–water two-phase flow status monitoring and evolution analysis [191]	Fluid dynamics and process monitoring, specifically focusing on industrial process monitoring and two-phase flow analysis in gas-water systems	Machine Learning Models	2024
				Problem Type: Classification and regression. Approach: Multi-task learning (MTL) framework to jointly mine sensor data and fluid properties, supplemented by SEDA for latent properties and Rank SVM for monitoring flow status. Topic: Real-time monitoring and analysis of gas-water two-phase flow statuses. Models Used: Multi-Task Learning (MTL), Sparse Exponential Discriminant Analysis (SEDA), Rank SVM. Year of Publication: 2024
88		X	X	Supervised image classification model for coral bleaching detection using a bi-temporal sentinel-2 image stack [192]	Coastal ecosystem monitoring and environmental change assessment, specifically focusing on coral reef monitoring and climate change impact on marine ecosystems.	Tree-Based Models	2024
				Problem Type: Classification and detection/segmentation. Approach: Using Sentinel-2 satellite images for pre- and post-bleaching events to apply a supervised Random Forest classifier for change detection. Topic: Detection of coral bleaching in Sekisei Lagoon, Japan. Models Used: Random Forest. Year of Publication: 2024
89	X			Enhancing estuary salinity prediction: A Machine Learning and Deep Learning based approach [193]	Water quality prediction and environmental monitoring, specifically focusing on salt wedge intrusion in estuarine environments and salinity prediction for sustainable ecosystem management	Tree-Based Models Neural Network Model Deep Learning Model	2024
				Problem Type: Regression. Approach: Machine learning and deep learning models were trained on data collected from the Po River estuary to predict salinity levels and compare them with a physics-based model. Topic: Estuary salinity prediction to address salt wedge intrusion in the Po River estuary. Models Used: Random Forest, Least-Squares Boosting, ANN, LSTM. Year of Publication: 2024
90	X	X		A settling velocity formula for irregular shaped microplastic fragments based on new shape factor: Influence of secondary motions [194]	Microplastic pollution Water Pollution Environmental Modeling	Numerical/Mathematical Modeling (Primary method) Deep Learning Models Tree-based Machine Learning Models Regression-based Machine Learning Models	2024
				Problem Type: Primarily Regression; Secondary Classification Approach: Developed a new explicit mathematical formula based on numerical simulations and physical experiments to predict microplastic settling velocities. Topic: The research focuses on environmental pollution by microplastics in water environments, specifically examining the settling velocities of irregular-shaped microplastic fragments. Understanding settling velocity is crucial for predicting microplastic migration, distribution, and environmental fate. Models Used: Mathematical and numerical modeling; Machine Learning methods (SVR, RF, CNN) for comparison only Year of Publication: 2024
91	X	X		On the use of hydrodynamic modelling and random forest classifiers for the prediction of hypoxia in coastal lagoons [195]	Water Pollution Coastal Ecosystems Management Environmental Modeling Climate Change Impacts	Machine Learning Models Numerical Modeling Interpretability Techniques	2024
				Problem Type: Classification (dissolved oxygen levels classification), Forecasting (hypoxic events) Approach: Integrated use of hydrodynamic numerical models and Random Forest classifiers, validated with field observations Topic: Prediction and forecasting of hypoxia in eutrophic coastal lagoons Models Used: Random Forest classifiers, SHAP analysis Year of Publication: 2024
92	X	X		Performances of three representative snow depth products originated from passive microwave sensors over the Mongolian Plateau [196]	Snow Monitoring Climate and Environmental Management Remote Sensing Applications Disaster Prevention and Mitigation Water Resource Management	Machine Learning Models Empirical and Semi-Empirical Remote Sensing Models Assimilation and Bayesian Models Hybrid Models	2024
				Problem Type: Regression (continuous snow depth prediction); Detection (presence/absence of snow). Approach: Comparative validation of snow depth products derived from passive microwave sensors, involving Random Forest machine learning integrated with physical snow models. Topic: Evaluation and validation of snow depth retrieval methods on the Mongolian Plateau. Models Used: Machine Learning (Random Forest), empirical satellite-based algorithms, Bayesian assimilation algorithms integrated with in situ measurements. Year of Publication: 2024
93	X			Comparing modeled predictions of coral reef diversity along a latitudinal gradient in Mozambique [197]	Marine Biodiversity Conservation Environmental Modeling Climate Change Impact Coastal Ecosystems Management	Tree-based Machine Learning Models	2024
				Problem Type: Primarily Regression (continuous prediction of biodiversity metrics: number of fish species and coral taxa). Approach: Boosted Regression Tree models trained on regional (WIO) and local (Mozambique-specific) environmental data. Topic: Predictive modeling of coral reef biodiversity along the Mozambican coastline. Models Used: Boosted Regression Trees (machine learning). Year of Publication: 2024
94	X			Construction of a Real-Time Forecast Model with Deep Learning Techniques for Coastal Engineering and Processes: Nested in a Basin Scale Suite of Models [198]	Coastal and Estuarine Hydrodynamics Environmental Modeling Climate Change and Coastal Risk Management	Deep Learning Model	2024
				Problem Type: Regression (real-time forecasting of water level, salinity, temperature, and velocity fields) Approach: Deep learning-based approach (CNN) integrated with numerical hydrodynamic models (Delft3D) Topic: Coastal hydrodynamic forecasting for coastal engineering and environmental management Models Used: Deep Learning Models (CNN), Numerical Hydrodynamic models (Delft3D) Year of Publication: 2024
95	X			Long-term water quality assessment in coastal and inland waters: An ensemble machine-learning approach using satellite data [199]	Water Pollution Monitoring Environmental and Resource Management Remote Sensing Applications Harmful Algal Bloom (HAB) Management Climate Change and Ecosystem Health	Machine Learning Models	2024
				Problem Type: Regression (estimation of Chl-a, aCDOM(440), turbidity); implicit detection (Harmful Algal Blooms indirectly via water quality estimation) Approach: Ensemble ML models trained on optical remote sensing data from satellites (MODIS-Aqua, Sentinel-2 MSI, PlanetScope), validated with in situ measurements Topic: Water quality assessment using satellite remote sensing in coastal and inland waters Models Used: Ensemble ML models (SVR, RF, GBDT, XGBoost, AdaBoost, Extra Trees, LightGBM, CatBoost) Year of Publication: 2024
96	X			Kolmogorov-Zurbenko filter coupled with machine learning to reveal multiple drivers of surface ozone pollution in China from 2015 to 2022 [200]	Air Pollution Environmental Health Climate Change and Meteorological Impact Environmental Modeling and Management	Machine Learning Models Deep Learning Models Explainability Techniques Hybrid Models	2024
				Problem Type: Regression (Predicting continuous ozone levels using meteorological and anthropogenic factors). Approach: Integration of Machine Learning with KZ filtering and SHAP for interpretability to separate meteorological effects from anthropogenic-driven ozone trends. Topic: Identifying and quantifying drivers of ozone pollution in China and revealing anthropogenic-driven ozone trends. Models Used: Decision Tree, Random Forest, XGBoost, ANN, MLP, coupled with SHAP interpretability. (XGBoost selected as primary model) Year of Publication: 2024

,

Table A2. Occurrences of types of AI Tasks.

Types AI Tasks	Occurrences
Regression	56
Classification	46
Clustering	25

,

Table A3. Occurrences of Topics.

Topics	Occurrences
Marine Ecology	9
Environmental Monitoring	21
Marine Biology	1
Environmental Genomics	1
Aquaculture Management	1
Water quality	34
Air quality	2
Palaeoclimatology	1
Land Use and Coastal Ecosystem Mapping	29
Flow Dynamics	7
Wetland Monitoring (Conservation)	5
Eutrophication Management	2
Groundwater Monitoring	1
Biodiversity and Conservation	4
Urban Expansion	1
Habitat Monitoring	2
Habitat Classification	1
Water resources management	4
Urban Climate and Air Temperature	1
Climate Change and Water Resources	8
Behavioral Ecology	1
Sustainable Agriculture and Ecosystem Services	1
Structural Engineering	1
Reservoir Characterization and Geochemistry	1
Photocatalysis	1
Traceability and food safety	1
Seismic Monitoring	1
Climate modeling	1

,

Table A4. Occurrences of AI Model Types.

Model Types	Occurrences
Machine Learning Models	146
Deep Learning Models	24
Statistical and Linear Models	14
Domain-Specific Models	7
Algorithmic Models	3
Decision and Explainability Models	2
Graph and Profile-Based Models	2

,

Table A5. Temporal distribution of publications.

Year	Occurrences
2024	31
2023	33
2022	20
2021	20
2020	6
2019	3
2018	2
2017	2
2016	2
2015	0
2014	2
2013	0
2012	0
2011	1
2010	1
2009	0
2008	0
2007	0
2006	0
2005	0
2004	1

and

Table A6. Occurrences of temporal trend and AI task type combination.

	Regression	Classification	Clustering
2004	1	0	0
2005	0	0	0
2006	0	0	0
2007	0	0	0
2008	0	0	0
2009	0	0	0
2010	0	1	0
2011	1	0	0
2012	0	0	0
2013	0	0	0
2014	1	0	1
2015	0	0	0
2016	2	0	0
2017	1	1	0
2018	1	1	0
2019	0	2	1
2020	4	1	1
2021	7	11	2
2022	10	3	7
2023	11	14	8
2024	17	12	5
total	56	46	25

; Figure 2, Figure 3, Figure 4, Figure 5 and Figure 6).

4.8. PRISMA-Style Flow Diagram

Figure 2 reports the study selection process using the official PRISMA 2020 flow diagram template for new systematic reviews (databases/registers). The diagram explicitly documents the number of records identified, removed prior to screening, screened, excluded, assessed for eligibility, and included, together with the reasons for full-text exclusions, thereby ensuring transparency and reproducibility of the selection process.

5. Results

Under this systematic review, the final evidence base comprised 96 peer-reviewed journal articles. To facilitate interpretation of the descriptive synthesis, Appendix A provides four complementary summary tables. Table A1 reports the study inventory used for coding (Paper_ID) and links each included record to the analytical categories adopted in this review. Table A2 summarises the frequency of AI task types (regression, classification, and clustering), noting that multiple tasks may co-occur within the same study; therefore, task occurrences can exceed the number of included articles. Table A3 reports the distribution of application topics addressed by the reviewed literature, highlighting the dominant thematic areas investigated in transitional-water contexts. Finally, Table A4 summarises the AI model families adopted across studies (e.g., ML, DL, and statistical/linear approaches), thereby providing a compact overview of methodological prevalence prior to the trend analyses reported in the following subsections.

5.1. Classification of Types of AI Tasks

AI tasks can be broadly categorized based on their objectives and learning methods. Among these, three primary types—Regression, Classification, and Clustering—are particularly relevant in ecological and environmental research (Table A2, Figure 2). These methods provide essential analytical tools for predicting relationships between variables, categorizing ecological data, and identifying patterns within complex datasets.

Across the reviewed corpus, regression represents the most frequently reported task (56 occurrences), followed by classification (46) and clustering (25) (Table A2). The total exceeds the number of included studies (n = 96) because a substantial fraction of papers reported more than one analytical objective, for example, coupling regression-based estimation (e.g., continuous water-quality variables) with classification (e.g., trophic status categories) or using clustering as an exploratory step to define ecological typologies prior to supervised modelling.

• Regression

Regression is a fundamental AI task that involves predicting a continuous output variable based on input features. It is widely used in ecological and environmental sciences for predicting temperature changes, pollution levels, and species population trends. Techniques such as linear regression, support vector regression, and neural networks have been applied in various fields, including climate modeling and biodiversity assessments [81]. AI-enhanced regression models improve predictive accuracy by integrating large datasets, making them particularly useful for forecasting ecological dynamics [55].

• Classification

Classification tasks involve assigning predefined labels to data points. In ecological AI applications, classification is often used for species identification, habitat categorization, and pollution detection. Traditional methods such as decision trees, RF models, and SVMs have been supplemented by DL models, which can achieve high classification accuracy in image recognition and remote sensing applications [58]. Recent advancements in hybrid AI-human classification models have further improved performance by combining automated predictions with expert validation [82].

• Clustering

Clustering is an unsupervised learning task that groups similar data points together. It is widely used in ecological modeling to identify environmental patterns, group species with similar characteristics, and analyze habitat distributions. Common clustering algorithms include k-means, hierarchical clustering, and density-based clustering (DBSCAN). In AI-driven ecological research, clustering has been instrumental in defining ecosystem boundaries and detecting emerging environmental threats [55].

• Other AI Tasks and Their Exclusion from This Study

Beyond the three primary categories explored in this study, other AI methodologies are commonly used in various scientific domains:

• Reinforcement Learning, which optimizes decision-making in dynamic environments but requires simulated scenarios beyond the scope of this research.
• Anomaly Detection, which is valuable for identifying rare ecological events, but is less applicable to structured datasets such as those used in this study.
• Generative AI, which synthesizes new data, is primarily beneficial for model training augmentation rather than direct ecological analysis [83].

This study focuses on Regression, Classification, and Clustering due to their well-established applications in ecological modeling and environmental analysis. These methodologies offer powerful tools for predicting environmental changes, classifying ecological features, and detecting patterns in complex datasets. While other AI approaches have their merits, the selected methods provide the most direct, interpretable, and actionable insights within the scope of this research.

• Findings on AI Task Type Distribution

The analysis of 96 studies revealed that Regression tasks are the most prevalent, followed by Classification and Clustering. Figure 2 illustrates the distribution of these AI task types, demonstrating their respective applications across various ecological domains.

• Regression (44.1%)—Used primarily for environmental forecasting, such as predicting pollution levels, climate trends, and species population dynamics.
• Classification (36.2%)—Applied extensively in species identification, habitat categorization, and environmental monitoring, leveraging decision trees, neural networks, and SVMs.
• Clustering (19.7%)—Used for grouping ecological data, identifying environmental patterns, and defining ecosystem boundaries through unsupervised learning techniques.

These results highlight the dominance of predictive and classification-based AI models in ecological research while showcasing the emerging role of clustering techniques in defining ecological relationships.

The findings reinforce the importance of regression-based predictions for ecological forecasting, classification for species and habitat identification, and clustering for uncovering hidden patterns in environmental datasets. Future advancements may further integrate these approaches to enhance AI-driven conservation and monitoring efforts.

5.2. Research Topics

The application of AI in ecological and environmental sciences spans a wide range of research topics. The following section categorizes the key research domains where AI techniques have been extensively applied, highlighting the primary challenges and opportunities they address (Table A3, Figure 3).

• Water Quality Monitoring

AI has been instrumental in assessing and predicting water quality parameters such as pollutant concentrations, nutrient levels, and turbidity. These applications play a crucial role in ecosystem health monitoring, sustainable water resource management, and public health protection [84]. Machine learning models, including Random Forest (RF) and ANN models, have been applied to detect anomalies and forecast changes in water quality over time [85].

• Land Use and Coastal Ecosystem Mapping

AI-driven models contribute significantly to land use classification and coastal ecosystem mapping, which are essential for tracking habitat distribution and assessing environmental changes due to anthropogenic pressures [86]. Remote sensing data, coupled with DL techniques such as CNN models, enhance the accuracy of these mappings [87]. These methodologies support conservation efforts and urban planning strategies by providing high-resolution spatial analyses.

• Marine Ecology and Environmental Monitoring

Machine learning and DL techniques are widely used to model species distribution, track biodiversity patterns, and monitor ecosystem changes over time [88]. AI-based predictive models enable the assessment of marine ecosystem resilience and inform policy decisions on marine conservation. Data-driven AI models also facilitate the real-time detection of harmful algal blooms and other environmental hazards [89].

• Climate Change and Ecosystem Responses

AI-driven climate models are employed to predict the impact of climate change on ecosystems. These models integrate large-scale climate data with ecological variables to project species adaptation, habitat shifts, and ecosystem resilience [90]. Hybrid AI models combining statistical and ML approaches offer improved forecasting accuracy for climate-related ecological changes [91].

• Other Specialized Topics

AI applications extend to niche areas such as environmental genomics, aquaculture management, and behavioral ecology [92]. AI-powered genomics helps in biodiversity conservation through genetic sequencing and evolutionary studies, whereas ML models optimize aquaculture by predicting species growth rates and monitoring fish health. Behavioral ecology studies leverage AI to analyze species interactions and movement patterns using tracking data [93].

• Findings on Research Topics

By structuring AI research topics in this manner, we provide a framework that reflects the fundamental environmental questions AI aims to address. This classification aids researchers and policymakers in navigating the literature and underscores the unique challenges and opportunities associated with each domain.

5.3. Distribution of AI Model Types

AI models used in ecological and environmental sciences exhibit significant methodological diversity, reflecting the complexity of the studied ecosystems. In this review, AI models were categorized into three primary types: ML Models, DL Models, and Statistical and Linear Models. This classification aims to highlight their applicability, strengths, and limitations in environmental monitoring and ecological research (Table A4, Figure 4).

• ML Models

ML models, including RF, SVMs, and ensemble methods, dominate AI applications in ecology due to their robustness in handling large datasets and uncovering complex patterns [94]. These models are widely applied in species distribution modeling, habitat classification, and water quality prediction. Studies have demonstrated that RF and SVM achieve high accuracy in predictive ecological models by effectively leveraging multi-dimensional datasets [95].

Importantly, the prevalence of ML models in the reviewed literature reflects not only their predictive performance but also their alignment with the typical constraints of transitional-water datasets. These datasets often combine moderate sample sizes with high-dimensional predictor spaces, irregular temporal sampling, and strong correlations among environmental variables. Under such conditions, ML approaches that rely on flexible function approximation and weak distributional assumptions offer practical advantages over classical parametric models.

In contrast, DL models are preferentially adopted when the dominant data modality is unstructured or weakly structured (e.g., imagery, acoustic signals, dense time series), whereas traditional statistical models remain relevant when data availability, interpretability, and hypothesis-driven inference are prioritized.

• DL Models

Deep learning (DL) approaches, such as CNN models and RNN models, have revolutionized AI applications in ecology due to their superior performance in analyzing unstructured and high-dimensional data, including images, spatial maps, and time-series data. These models have been instrumental in remote sensing applications, biodiversity assessment, and marine ecosystem monitoring [96]. CNN models, in particular, have enhanced the classification accuracy of land cover and ecosystem mapping by learning intricate spatial features [97].

• Statistical and Linear Models

Traditional statistical models, including linear regression, generalized linear models (GLMs), and Bayesian models, continue to be relevant in ecological research. These models provide baseline predictions, hypothesis testing frameworks, and interpretable analytical tools, making them essential for understanding the behavior of more complex AI models. Despite the rise of DL techniques, statistical models remain integral to many ecological studies, particularly in climate modeling and population dynamics [98].

• Other Specialized AI Models

While ML, DL, and statistical models dominate AI research, other model categories exist, focusing on domain-specific ecological applications, algorithmic innovations, and simple explanations [99]. XAI models are particularly crucial in ecological research, ensuring transparency, model interpretability, and trust in AI-generated predictions [95].

• AI Model Type Distribution and Trends

The reviewed literature indicates a clear prevalence of ML models, followed by DL and statistical methods, with specialized approaches emerging as niche areas [53]. This trend highlights the ongoing refinement of AI methodologies to ensure accuracy, interpretability, and ecological relevance in environmental research.

• Findings on AI Task Type Distribution

By structuring the classification of AI models in this manner, we provide a systematic framework for researchers and policymakers, enabling a clearer understanding of how different AI methodologies align with specific ecological challenges. These findings emphasize the importance of selecting appropriate AI models based on data characteristics, research objectives, and the need for interpretability in ecological decision-making.

5.4. Temporal Distribution of Publications

The evolution of research in AI applied to ecological and environmental sciences has followed a significant trajectory, reflecting both technological advancements and shifting research priorities. Analyzing the temporal distribution of publications allows us to identify key trends in AI applications and their adoption in environmental monitoring and conservation (Table A5, Figure 5).

• Trends in AI Research Over Time

The temporal analysis of AI-related publications indicates a notable increase in research output starting from 2018, with a particularly sharp rise after 2020. This surge aligns with several factors, including:

• Advancements in computational power, particularly the accessibility of GPUs and cloud computing, which have facilitated the adoption of ML and DL approaches.
• An increase in high-quality environmental datasets, made available through remote sensing satellites and sensor networks, enabling the training of more sophisticated AI models.
• A growing emphasis on climate change research, leading to increased funding and international collaboration on AI-driven ecological monitoring.

• Key Periods of Growth and Innovation

The review of publication trends highlights distinct periods of accelerated AI adoption:

• Before 2015: the use of AI in ecological research was minimal, with a strong reliance on traditional statistical models and early ML methods like Decision Trees and Regression models.
• 2015–2019: A steady rise in publications, with increased experimentation using SVMs and RF models for species distribution modeling and environmental classification.
• 2020–2024: A substantial increase in AI publications, driven by the rise of DL methodologies such as CNN models for remote sensing, as well as hybrid AI approaches integrating ML with ecological models.

The temporal distribution of AI research thus provides valuable insights into the maturation of the field, helping to contextualize both past achievements and future directions.

From an ecological data perspective, the post-2020 increase in deep-learning studies is consistent with the growing availability of high-resolution remote-sensing products and continuous sensor records, for which convolutional and sequence-based architectures are often methodologically appropriate, provided that validation accounts for spatio-temporal dependence and domain shift across sites and seasons.

• Impact of AI on Environmental Research Priorities

The increasing number of AI-driven ecological studies over the past decade suggests a shift in research priorities:

• Early AI applications (before 2015) were primarily focused on species monitoring and environmental forecasting, relying on structured datasets with limited computational requirements.
• Recent studies (2020–2024) have expanded to real-time monitoring, anomaly detection, and predictive modeling, enabled by DL and big data analytics.
• A parallel trend in XAI has emerged to enhance model interpretability, addressing concerns about the transparency and reliability of AI models in ecological decision-making.

These trends reflect both technological progress and an increasing urgency in tackling climate-related environmental challenges, positioning AI as a critical tool in ecological research and conservation strategies.

The analysis of temporal publication trends underscores the accelerated integration of AI in environmental sciences, with a shift from traditional statistical models to more sophisticated ML and DL approaches. This growth reflects improvements in data availability, computational resources, and global environmental concerns, shaping the future of AI-driven ecological research.

5.5. Temporal Trends and AI Task Types Combination

The evolution of AI task types in ecological and environmental research follows a distinct temporal pattern, influenced by technological advancements, increasing computational resources, and the availability of high-resolution datasets. By analyzing publication years alongside AI task types—Regression, Classification, and Classification, this section highlights the shifts in modeling approaches and their alignment with research priorities over time (Table A6, Figure 6).

• Historical Evolution of AI Task Types

Before 2015, AI use in environmental sciences was largely centered on regression models, which played a key role in continuous environmental monitoring, species distribution estimation, and climate projections. Methods like linear regression, support vector regression (SVR), and Gaussian processes were commonly employed for their clarity and effectiveness in handling structured ecological data.

Between 2015 and 2019, classification models gained prominence, driven by the growing availability of labeled datasets and the adoption of tree-based methods, RF models, and SVMs. These methods were particularly useful for land cover classification, biodiversity assessments, and pollution detection.

From 2020 onwards, the increasing availability of high-resolution remote sensing data and long-term environmental monitoring datasets coincided with a rapid adoption of deep learning models, primarily for supervised tasks such as classification, regression, segmentation, and time-series forecasting. While unsupervised methods, including clustering, continue to play a role in exploratory ecological analyses, deep learning architectures are most commonly employed either in supervised settings or as feature extraction mechanisms, where learned representations are subsequently analyzed using conventional clustering algorithms.

Therefore, the observed use of clustering in recent studies should not be interpreted as a direct consequence of deep learning adoption, but rather as part of hybrid analytical pipelines combining representation learning with traditional unsupervised techniques.

• Factors Influencing the Shift in AI Task Types

The increasing adoption of classification and segmentation tasks in recent years can be attributed to:

• Enhanced computational capabilities, including cloud computing and GPUs, have facilitated the training of complex models.
• The growing availability of high-resolution environmental data, especially from satellite imagery and remote sensing, has expanded analytical possibilities.
• The demand for real-time ecological monitoring has increased, necessitating models that can detect anomalies and identify environmental patterns.
• Escalating concerns over climate change and biodiversity loss have fueled the adoption of advanced AI-driven analytical approaches.

• AI Task Type Combinations Over Time

• Before 2015: Regression models were dominant, supporting species distribution and pollution forecasting.
• 2015–2019: Classification tasks emerged, used in species identification and land use mapping.
• 2020–2024: Increased use of clustering methods was observed, primarily as part of exploratory analyses or hybrid workflows combining deep-learning-based feature extraction with traditional unsupervised techniques, rather than as a direct consequence of deep learning adoption.

This temporal shift in AI applications underscores a progressive transition toward higher-dimensional, real-time, and data-driven ecological monitoring, reflecting both methodological advancements and increasing environmental research needs.

By integrating temporal trends with AI task types, this analysis provides a comprehensive timeline of AI evolution in ecological research (Figure 7). The observed patterns highlight how AI methodologies have expanded from simple predictive models (regression) to sophisticated classification and segmentation approaches, driven by technological innovation and environmental urgency. Future research is expected to further refine detection models and enhance explainability, ensuring robust and interpretable AI-driven ecological solutions.

6. Discussions

This review provides a comprehensive overview of the application of AI—specifically ML and DL—in the study of transitional water ecosystems. The results highlight a rapid evolution in AI methodologies, evidenced by the significant temporal increase in publications and the diversification of AI tasks and model approaches.

6.1. Shift in AI Model Adoption

A notable trend in recent research is the shift from predominantly regression-based models to a more diverse use of classification and clustering techniques. This evolution is driven by:

• Advances in computational power and DL, allowing for more complex, high-dimensional analyses.
• The increasing complexity of ecological research, necessitating sophisticated analytical approaches.
• Greater adoption of clustering models, particularly in remote sensing, habitat mapping, and pollution assessment.

For example, CNN models have significantly improved habitat classification and species monitoring by enhancing spatial accuracy and detection capabilities.

6.2. Trends in AI Model Type Distribution

The distribution of AI model types further emphasizes this evolution:

• Traditional ML methods, such as RF models, SVMs, and Gradient Boosting, remain widely used due to their robustness and their perceived transparency, although most high-performing approaches lack intrinsic interpretability in ecological applications.
• The rising adoption of DL models, particularly CNNs and RNN models, suggests that researchers are increasingly leveraging high-dimensional data, such as satellite imagery and sensor-based time-series data.

This trend reflects an increasing focus on automated feature extraction and real-time monitoring, which is crucial for climate adaptation strategies and biodiversity assessments.

Although traditional ML models are sometimes preferred for their perceived transparency, most high-performing approaches used in recent ecological studies lack intrinsic interpretability, reinforcing the need to distinguish clearly between model transparency and post hoc explainability.

6.3. AI and Global Environmental Concerns

The temporal trends in AI adoption also highlight the connection between AI advancements and growing global environmental concerns:

• Climate change and habitat degradation have catalyzed innovation in AI applications, particularly in predictive modeling and ecosystem monitoring.
• The expanding volume of AI-driven ecological publications not only marks technological advancements but also underscores the increasing urgency for effective environmental monitoring and management strategies.

These trends indicate a paradigm shift in ecological research, where AI is no longer just a supplementary tool but a critical component in policy-driven conservation efforts and environmental sustainability.

6.4. Challenges and Future Directions

Despite the promising advancements, several challenges remain:

• Model Interpretability—The increasing complexity of AI models necessitates XAI techniques to ensure transparency in ecological decision-making.
• Data Quality and Availability—AI models require large-scale, high-quality datasets, which remain limited in some ecological contexts.
• Computational Constraints—While AI models provide significant advantages, their high computational costs and energy consumption pose sustainability concerns.

Future research should focus on:

• Enhancing XAI methods to improve model transparency and stakeholder trust.
• Developing standardized AI frameworks to facilitate cross-study comparability and model reproducibility.
• Integrating AI with ecological domain expertise, ensuring that AI-generated insights align with real-world conservation priorities.

The integration of diverse AI methodologies into ecological research has transformed our ability to analyze complex, large-scale environmental datasets. While these advances offer promising avenues for improved ecosystem monitoring and conservation, they also highlight the need for interdisciplinary collaboration, improved model interpretability, and better data infrastructure. Future research should prioritize transparency, scalability, and ecological relevance, ensuring that AI serves as a reliable and sustainable tool in environmental management.

7. Conclusions and Recommendations

The adoption of AI in environmental sciences has revolutionized ecological monitoring, conservation efforts, and sustainable management. This review explores the rapid advancement of AI applications in these fields, highlighting the transition from traditional statistical techniques to more sophisticated ML and DL approaches. AI now plays a crucial role in addressing environmental challenges such as biodiversity preservation, pollution detection, climate prediction, and ecosystem management [88].

7.1. Key Findings and Implications

• Advancements in AI-driven Environmental Monitoring

AI-based models, particularly CNN models and ensemble ML techniques, have improved the accuracy and efficiency of ecosystem monitoring by analyzing vast datasets from satellites, drones, and sensor networks [87].

• AI’s Role in Conservation and Sustainable Practices

AI is widely employed in species identification, habitat assessment, and predictive conservation strategies. AI models utilizing acoustic recordings, camera trap imagery, and eDNA analysis have enhanced biodiversity tracking and anti-poaching initiatives [100].

• Climate Change Mitigation and AI

AI models contribute significantly to climate risk assessment, predictive environmental modeling, and mitigation strategies by analyzing global climate patterns and forecasting extreme weather events [101].

• Challenges in AI-driven Environmental Research

  a.

  Data Availability and Bias: Many AI models depend on high-quality datasets, yet access to consistent, standardized ecological data remains a challenge [86].

  b.

  Computational Costs: Deep learning models require significant computational resources, leading to concerns over energy consumption and sustainability [84].

  c.

  Model Interpretability: The need for XAI is growing, as complex models often lack transparency, making it difficult for policymakers and conservationists to rely on AI-based decisions [102].

7.2. Recommendations for Future Research and Policy

• Enhancing AI Transparency and Explainability

The development of XAI should be prioritized to ensure that AI models in ecological and conservation research are interpretable, trustworthy, and actionable

• Improving Data Accessibility and Standardization

Policymakers and researchers should collaborate to create open-access environmental datasets, improving model accuracy and comparability across studies [36].

• Integrating AI with Traditional Ecological Knowledge (TEK)

Future AI models should incorporate indigenous and local ecological knowledge to improve conservation efforts and align predictions with real-world observations [103].

• Ensuring Ethical and Sustainable AI Use

Governments and institutions should enforce ethical AI guidelines, including considerations of fairness, accountability, and sustainability in environmental AI applications [93].

7.3. Conclusions

AI has emerged as a transformative tool in environmental research, enhancing ecological modeling, conservation efforts, and climate change mitigation. However, its effectiveness depends on ethical deployment, cross-disciplinary collaboration, and strong data infrastructure. Tackling issues such as data bias, model interpretability, and computational sustainability is essential to ensuring AI’s long-term contribution to environmental protection. Future research should emphasize integrating AI with traditional conservation strategies while advancing XAI and data-sharing frameworks to improve ecological monitoring and climate resilience.

A responsible and collaborative AI ecosystem—driven by researchers, policymakers, and conservationists—can maximize AI’s benefits while aligning with global sustainability objectives [104].