In coastal engineering, data can come in different forms (Huang et al., 2015) and can be classified into different types based on their identity, format, and structure. Some examples of coastal data types include:

(1) Numeric data (Timmermans et al., 2020), which includes measurements of various physical parameters such as water level, wave height, current velocity, sediment concentration. Such data are typically collected using instruments such as tide gauges, wave gauges, current meters, and sediment samplers. For example, time-series data such as ocean temperature records, sea level measurements, and storm surge data represented by a sequence of observations or measurements taken at regular intervals over time.

The following are the most well-known methods for collecting the data mentioned above: field observations, remote sensing measurements, experimental studies, numerical and mathematical models. Both the availability of equipment and the objective of the study influence the selection of the data collection medium (Prata et al., 2019).

Data collection within the realm of marine sciences principally relies on three distinctive methods: in-situ observations, remote sensing techniques, and the use of mathematical and numerical models, as outlined by Verwega et al. (2021). In-situ data collection encompasses ship-based measurements, the deployment of moorings, gliders, autonomous underwater vehicles, drifters and floats, the use of sea-floor optic cables, and laboratory analyses. Field observations remain essential for the collection of real-world data on coastal processes, such as wave heights and tidal levels. In-situ instruments are highly accurate with proper maintenance but may have low-time frequency data for large areas. They offer historical climate trend insights not available from remote sensing and are less affected by atmospheric conditions. These observations serve to validate numerical models that simulate coastal processes and predict the behavior of the coastal system, including wave patterns, tidal currents, and shoreline evolution.

Remote sensing involves acquiring data on coastal topography, bathymetry, and other significant parameters through satellite and airborne platforms. Remote sensing technologies are divided into three categories: satellite, ground-based, and drones (Elsayed et al., 2021). The data thus collected enable the generation of high-resolution coastal environmental maps. Although satellites are powerful tools, they face limitations in obtaining high-resolution regional-scale imagery. Clouds can hinder data capture, and high-resolution imagery can be challenging to interpret (Elsayed et al., 2021). A combination of satellite- and ground-based remote sensing and drones could be effective in future marine engineering evaluations. Economically, combining these tools may be comparable to in-situ techniques in terms of overall cost. Such technology could enable rapid, high-resolution water condition assessments and enhance our understanding of water resource processes. Mathematical and numerical models generate data by simulating real-life systems or processes using mathematical equations and algorithms (Xie and Arkin, 1996). They provide the capability to extend observational data, even to the point of simulating future climate scenarios (Eyring et al., 2016). Nonetheless, it is crucial to understand that these models only approximate real-world scenarios and can encompass spatial and temporal scales that exceed the scope of observational data (Matthes et al., 2020). The outputs from these models are typically available on a unique grid, contingent on the specific simulation. For instance, climate models customarily provide a four-dimensional space-time grid. Consequently, the comparison of model outputs with measurements invariably necessitates interpolation or data aggregation. Table 1 provides a detailed summary of the advantages and disadvantages associated with these diverse data collection methodologies.

Several factors can influence the quality of observational data. These include inaccuracies in the instruments, malfunctions of the equipment, disruptions from external sources, mistakes during data conversion, communication mishaps, and significant unforeseen errors (Yu et al., 2022). Such anomalies can pose major threats to operational functionality, downstream operations, system resilience, and cleaner production (Ba-Alawi et al., 2021). Therefore, these should be detected promptly and their data rectified to ensure more realistic measurements.

Anomaly detection methods are generally categorized into various types (see Figure 2) such as Statistical Methods, that utilize the properties of the underlying data distribution to identify anomalies (Chandola et al., 2009); Distance-based Methods, which calculate the distance between data points and identify the outliers based on a certain distance threshold (Ramaswamy et al., 2000); Density-based Methods, which estimate the density of data points and identify outliers as those points that reside in low-density regions (Ester et al., 1996); Machine Learning-based Methods, which employ supervised, unsupervised, or semi-supervised ML algorithms to detect outliers (Pimentel et al., 2014); and Ensemble Methods, which combine multiple outlier detection algorithms to improve the overall performance (Zimek et al., 2012). The choice of method, or combination of methods for better results, depends on the nature of the data and the specific problem being addressed.

Mahmoodi and Ghassemi (2018) used outlier detection algorithms to improve wave height predictions, while Oehmcke et al. (2015) demonstrated the effectiveness of ML for identifying significant events in marine long-term data. Daranda and Dzemyda (2020) developed a method combining the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering algorithm and k-nearest neighbors analysis for detecting marine traffic anomalies. These studies highlight the potential of leveraging advanced algorithms and ML in marine data analysis and decision-making. This section aims to provide a survey of contemporary outlier detection techniques, comparing their motivations, advantages, and disadvantages. Outliers can significantly impact the results, which makes addressing or eliminating them before analysis and model development crucial.

Considering the learning algorithm, three main methodologies exist for outlier detection (Hodge and Austin, 2004): 1) unsupervised approach, which uses a learning technique to identify outliers without prior knowledge of the data. The data is treated as a static distribution, and the most distant points are flagged as potential outliers; 2) supervised classification method, which requires pre-labeled data. It allows for online classification, where the classifier continuously learns the model and classifies new data as normal or abnormal, and finally 3) semi-supervised recognition technique, which only learns the normal class, using pre-classified data. It can distinguish new data as normal or novel based on its proximity to the boundary of normality. The choice of an outlier detection method depends on the data type, the number of vectors and attributes, speed and accuracy requirements, and the ability to accurately identify outliers. The key factors in choosing a method are selecting an algorithm that can handle the data and defining a suitable neighborhood for the outlier.

Incorporating parameters that are not relevant can result in intricate models that pose significant challenges in interpretation and execution compared to the models developed using the most crucial parameters (Pourzangbar, 2012). That is the reason why the focus is placed on building ML models using the most crucial parameters. These parameters are not only essential for the model’s output, but also are unconnected with other input parameters. To derive the most important dimensions (parameters) in the input space, there are several methods including min/max autocorrelation factor analysis (MAFA), dynamic factor analysis (DFA), Least Absolute Shrinkage and Selection Operator (LASSO), Independent Component Analysis (ICA), multicollinearity test and PCA. Table 2 summarizes some famous dimensionality reduction approaches used in marine engineering. The latter two methods are explained below.

Although numerous methods exist for DA, the majority of studies employ the Buckingham π Theorem to render the parameters dimensionless. Table 3 summarizes some of the studies used DA before feeding their ML models.

Normalizing data helps to ensure comparability by transforming it into a common scale, avoiding bias in statistical analyses and allowing for accurate and meaningful results by removing the impact of unit differences, especially when comparing data from different sources. Normalization plays a crucial role in efficient machine and deep learning by ensuring that large numerical inputs are processed effectively (Van Komen et al., 2022). The choice of normalization method depends on the specific requirements of the data and the problem being solved. Some of the famous methods for data normalization are summarized in Table 4.

In Table 4, the transformed data, referred to as x N , is obtained by normalizing the original data ( x i ) in a new range. The original data is contained within a vector, denoted as X , and its minimum and maximum values are represented as min X and max X , respectively. The chosen minimum and maximum values for the transformed range are N e w max and N e w min , which are typically set to zero and one, respectively. μ X is the mean of the data, and σ X is the standard deviation of the data.

Min-Max normalization is a technique used to rescale a feature to a specific range, usually between 0 and 1. However, to avoid having zero data in the model, an alternative approach is to expand the range to include values between 0.1 and 0.9. It is a commonly used method for transforming variables so that they are comparable, as it scales the data linearly to a specific range. Through this normalization process, the values in x N are transformed such that the minimum value of X is mapped to 0, the maximum value to 1, and intermediate values are mapped to corresponding values between 0 and 1. The Z-score normalization, also known as standardization, is a method of transforming data to a standard normal distribution with a mean of 0 and a standard deviation of 1. This normalization process rescales the data and centers it around the mean, allowing for easier comparison of values. It is commonly used in various fields, such as statistics, ML and data analysis. Sigmoid normalization uses a sigmoid function to transform the data, proving useful in instances where the data distribution is asymmetrical. The sigmoid function maps any input value to a value between 0 and 1 and it is commonly used in ML and ANN models to represent a probability or to rescale data. Additionally, the sigmoid function is differentiable, which makes it useful in optimization problems and backpropagation in neural networks.

In coastal phenomena, the relationship between inputs and outputs typically displays nonlinearity, but certain models, such as the M5 model tree, are unable to handle nonlinearity. To address this limitation, M5 models have been implemented using a logarithmic form for both inputs and outputs (i.e., the natural logarithm of inputs and outputs). This logarithmic form is more accurate than a linear formulation because it better captures the nonlinear nature of the contributing parameters (Pourzangbar et al., 2017a; Afsarian et al., 2018). Log scaling entails transforming data points through the application of a logarithmic function. The logarithm maps large values to smaller ones and vice versa, helping to make skewed data more symmetrical and manageable for analysis. The selection of a specific logarithmic function depends on the needs of the data and the analysis to be performed, such as log base 10, log base 2, or natural logarithm. Despite its advantages, normalization may result in a loss of interpretability, increased sensitivity to outliers (as seen in techniques like min-max scaling and z-score), loss of information, dependence on the entire dataset, impacts on categorical features, and varying sensitivity across algorithms.

Supervised ML presents a powerful approach, necessitating labeled data for model training. Its versatility permits its usage across a variety of applications, such as image and speech recognition, natural language processing, and predictive analytics. Common algorithms used in supervised learning encompass linear regression, logistic regression (LR), decision trees, random forests, support vector machines, and neural networks. A key advantage of supervised learning is its capacity to generate precise predictions for novel and unseen data (Jiang et al., 2020). However, it also has certain drawbacks, including the requirement for labeled data, the quality and quantity of the training data, and the potential for overfitting. Despite these challenges, supervised learning is seen as an essential method in ML and data science, demonstrating high accuracy and less computational time compared to physical models. Despite the inherent complexity of marine processes, supervised-based ML models have demonstrated benefits in understanding coastal phenomena, thereby finding extensive application in coastal engineering to drive innovative models and solve intricate problems (as summarized in Table 5). Supervised ML models have been employed to predict wave parameters like significant wave height and period, wave reflection and transmission coefficients (van Gent et al., 2007; Gandomi et al., 2020; Kuntoji et al., 2020), tide levels (Lee, 2004), ocean currents and wind files (James et al., 2018; Shamshirband et al., 2020), prediction of wind Characteristics under future Climate Change scenarios (Yeganeh-Bakhtiary et al., 2022), flood inundation using Gaussian process model (Donnelly et al., 2022) and breakwater stability number and wave overtopping discharge, among others. Various ML models, such as ANN and SVM, can be employed to do these predictions. ML models have also found application in morphological and morphodynamic predictions, including profile elevation, area, and length, based on parameters like wind speed, direction, wave height, and beach angle (Hashemi et al., 2010).

Unsupervised learning is a form of ML that functions without predefined labels or target outcomes (Bishop and Nasrabadi, 2006). Its main purpose is to independently discover patterns, structures, and relationships in data. Common applications include clustering, anomaly detection, and dimensionality reduction. Clustering groups similar data points, anomaly detection spotlights unusual patterns (as detailed in Section 2.3), and dimensionality reduction simplifies the number of features while preserving essential information (as seen in Section 2.4). Algorithms like k-means clustering, hierarchical clustering, PCA, and autoencoders are frequently used in unsupervised learning to identify patterns in data. While unsupervised learning can pose challenges due to the lack of a distinct optimization goal, it still holds a vital position in ML, contributing to advancements in fields such as computer vision, natural language processing, and recommendation systems. In the context of coastal engineering, k-means clustering can be used to classify centroid values for data like the maximum oceanic wind. Average centroid clustering can be obtained from both the previously chosen values and the currently selected clustering data (Baboo and Tajudin, 2013). PCA can be employed in coastal engineering to examine correlation matrices (Roseman et al., 2005) and pinpoint major changes in beach profiles and sand grain distributions (Tsujimoto et al., 2012). Moreover, PCA and hierarchical clustering can help characterize coastal plane shape and hydrodynamics. For instance, the form of arc-shaped coasts, largely influenced by geological structure, can be divided into four broad categories that reflect actual conditions using clustering (Scott et al., 2011). By identifying key data components, PCA can aid in elucidating the underlying patterns and structures of the data.

Reinforcement learning (RL) is a type of ML where a program, known as an agent, learns to perform tasks by getting feedback from its environment in the form of rewards or penalties (Rengarajan et al., 2022). The agent executes a series of decisions in a mutable environment, aiming to learn the optimal way (or policy) to maximize rewards over time. This process is typically structured as a Markov Decision Process (MDP), encompassing states, actions, transition functions, and reward functions. There are two main types of reinforcement learning algorithms: model-based and model-free. Model-based RL is like making a map to understand the surroundings. On the other hand, model-free RL does not make a map; it just figures out what to do based on where it is at the moment. So, model-based RL is more about planning ahead, while model-free RL is more about learning on the go (Plaat et al., 2023). Model-free methods, like Q-learning, do not need a model of the environment and calculate the expected total of future rewards for each possible action at each state using the so-called the Bellman equation. Q-learning has been used successfully in many different tasks, which is why it is one of the most commonly used model-free RL algorithms. In coastal engineering, RL can be used to develop control policies to reduce the risk of flooding (Bowes et al., 2021). Deep reinforcement learning (DRL), an advanced form of RL, can be used to control devices that convert wave energy, and has been found to work better than traditional control methods (Anderlini et al., 2020). DRL can also adjust itself to changes in system dynamics, allowing for control even when faults occur. Moreover, RL has been used to maximize the electricity produced by wave energy converters (Zou et al., 2022). In addition, a type of RL called multiagent reinforcement learning can simulate the social and economic effects of sea level rise. This can be a useful tool for planning scenarios, analyzing costs and benefits, and optimizing strategies to adapt to changes (Shuvo et al., 2022).

Table 4 summarizes the various ML learning approaches and their corresponding model types. Each model type utilizes a unique set of algorithms. For instance, in the case of classification tasks, ANNs or SVMs may be utilized. The final column of the table highlights the research studies focusing on each specific learning approach, targeting the investigation of a specific coastal process or event.

Predictive models, such as statistical, numerical, or ML models, play a vital role in marine and coastal engineering to safeguard structures from natural forces. Statistical models use past data to forecast future conditions, while numerical models simulate the event using mathematical equations and formulas. ML models, using artificial intelligence (AI), learn from past data for prediction purposes. Each model has its unique approach and is chosen based on data availability and specific project needs.

Various ML techniques have been implemented in the study of coastal and marine environments. Figure 3, sourced from Scopus, provides a visual representation of the percentage of published papers that used different ML methods since the year 2000. Upon reviewing this figure, it is evident that Principal Components Regression (PCR), Linear Model (LM), Regression Tree (RT), and ANN are the most frequently employed ML algorithms for analyzing coastal and marine phenomena. However, certain ML techniques, such as General Regression Neural Networks (GRNN), M5 model tree, Bayesian Model Averaging method (BMA), Generalized Boosted Regression (GBM), and Extreme Gradient Lift (Xgboost) have been applied less frequently in the investigation of coastal and marine events.

Figure 4 shows the application trend of different ML approaches for coastal and marine phenomena. Previously, techniques such as Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and RF were sparingly employed in diverse studies. However, there has been a significant increase in their use over the past 5 years, demonstrating a growing reliance on these methods in recent research.

Figure 5 illustrates the trend of various ML algorithms since 2008 in coastal and marine applications. Some methods are not frequently used which are colored in red (these low-important approaches are not reported in Figure 4).

• Developing hybrid models: the employment of combined and hybrid models has exhibited significant success, notably in addressing multifaceted issues. Eslaminezhad et al. (2022) advanced the efficiency of tree-structured machine learning models in determining the crucial parameters for forecasting flood susceptibility and constructing flood susceptibility maps, through the incorporation of the BPSO algorithm.