Exploring Recent Maritime Research on AIS-Based Ship Behavior Analysis and Modeling

Abstract

Automatic Identification System (AIS) data provide valuable insights into ship behavior, supporting maritime safety, situational awareness, and operational efficiency capabilities that are increasingly required for autonomous ship functions and harbor maneuvering assistance. This review synthesizes recent research on AIS-based ship behavior analysis and modeling published between 2022 and 2024 using a structured literature search and screening process informed by PRISMA principles. The review presents a five-stage workflow, spanning data processing, data analysis, knowledge extraction, modeling, and runtime applications with emphasis on how these stages contribute to perception, prediction, and decision support in automated navigation. Four dimensions are considered in data analysis, including statistical analysis, safety indicators, situational awareness, and anomaly detection. The modeling approaches are categorized into classification, regression, and optimization, highlighting current limitations such as data quality, algorithmic transparency, and real-time performance, while also assessing runtime feasibility for onboard or edge deployment. Three runtime application directions are identified: autonomous vessel functions, remote monitoring and control operations, and onboard decision-support tools, with numerous studies focusing on constrained waterways and port-approach scenarios. Future directions suggest integrating multi-source data and advancing machine learning models to improve robustness in complex traffic and harbor environments. By linking theoretical insights with practical onboard needs, this study provides guidance for developing intelligent, adaptive, and safety-enhancing maritime systems.

Keywords: machine learning; decision-support system; autonomous ship; knowledge extraction; AIS-based ship behavior modeling.

1. Introduction

Maritime transport is the backbone of global trade, accounting for over 80% of the world's merchandise by volume by sea [1]. Ensuring the safe and efficient operation of this vast network is critical. It supports economic stability, strengthens supply chain resilience, and advances environmental stewardship. However, maritime traffic is growing more complex. Rising vessel density, a diversity of ship types, and dynamic environmental conditions make monitoring and management more challenging than ever. The Automatic Identification System (AIS) is pivotal in maritime operations and addresses these challenges. It is a shipborne communication system. It continuously transmits vessel identity, position, course, speed, and heading via VHF radio signals [2]. The International Maritime Organization (IMO) mandates AIS for vessels over 300 gross tons.

AIS has since evolved from its origins in vessel recognition and security into a multifunctional data source that aids navigation, traffic regulation, maritime security, port logistics, and environmental monitoring. AIS data are particularly valuable due to their global coverage, public accessibility, and high temporal resolution. These attributes have positioned AIS as a primary data source for international studies on maritime traffic patterns and vessel operations [3]. Beyond its operational utility, AIS enhances situational awareness for crews, supports regularity enforcement, and enables real-time coordination of port activities.

One of the most active research domains enabled by AIS is ship behavior analysis, which examines the interaction between a vessel, its operating environment, and human decision-making. Ship behavior is shaped by a complex interplay of factors, including vessel design (e.g., hull geometry and propulsion systems), environmental conditions (e.g., wind, currents, and bathymetry), and navigational choices (e.g., speed adjustments and maneuvering) [4]. Understanding this behavior is essential for improving maritime safety and operational efficiency. Recent studies have leveraged AIS data to analyze and model ship behavior across various applications, including anomaly detection, risk assessment, traffic pattern analysis, and trajectory prediction [5,6]. Notable examples include AIS-based approaches for collision avoidance [7–9], behavior analysis [10], path planning [11,12], risk modeling [13,14], and deep learning-based trajectory prediction [15,16].

At a broader level, prior surveys have synthesized these directions from complementary perspectives: Tu et al. [17] organized AIS-enabled intelligent navigation into anomaly detection, route estimation, collision prediction, and path planning, while Xiao et al. [18] reviewed traffic-pattern mining and forecasting in maritime traffic service networks, emphasizing that traffic pattern knowledge is a foundation for forecasting and situation awareness. Collectively, these efforts demonstrate the potential of AIS to support safer, smarter, and more sustainable maritime operations.

Despite these advances, the existing literature remains fragmented across tasks and method families, often relying on different assumptions, datasets, and evaluation protocols, which limits comparability and hinders cumulative synthesis [19–21]. Although previous AIS surveys have provided valuable overviews of trajectory analysis, anomaly detection, traffic pattern mining, and related maritime applications, they have generally emphasized task-based categorization, algorithm inventories, or data characteristics rather than the interaction between methodological stages and deployment constraints. A focused review is therefore needed not only to consolidate the latest methodology, but also to provide a more integrative perspective that can inform the development of next-generation navigation and decision-support systems [22,23]. This paper addresses this need by providing a structured and comprehensive review of AIS-based methods for ship behavior analysis and modeling. It consolidates recent developments in the field, identifies key research challenges and gaps, and outlines future directions for advancing operationally feasible maritime analytics. Rather than proposing a new modeling method, this review contributes a workflow-based synthesis that links data processing, analysis, knowledge extraction, modeling, and runtime applications as interdependent stages. In addition, it proposes a modeling taxonomy that comparatively organizes major methodological paradigms in terms of their assumptions, outputs, and suitability under different data and operational conditions.

The remainder of this paper is organized as follows: Section 2 presents the literature review methodology, informed by PRISMA principles, and outlines the general workflow; Section 3 reviews data analysis approaches; Section 4 presents a taxonomy of modeling techniques; Section 5 discusses the main findings, limitations, and implications; and Section 6 concludes this paper.

2. Methodology

2.1. Literature Search Approach

This review adopted a structured literature search and screening process informed by PRISMA principles proposed by Moher et al. [24] in order to improve transparency in the identification, screening, and selection of relevant studies. Because the manuscript was not designed as a formal PRISMA 2020 systematic review, it is presented as a review of recent research rather than as a systematic review. The search targeted peer-reviewed journal articles and conference papers published between 2022 and 2024 that included the keyword "AIS" or "Automatic Identification System" in the title or abstract, combined with topic terms related to ship behavior (e.g., anomaly detection, collision risk, trajectory prediction, traffic monitoring, and decision support). Relevant studies were retrieved from three major bibliographic databases (Web of Science, Google Scholar, and IEEE Xplore) and complemented by two additional scholarly sources (ResearchGate and arXiv). The initial corpus was screened for relevance, language, publication type, and availability of full text. Studies examining the analysis, modeling, or decision-support applications using AIS data, as well as those providing methodological detail, were included. Materials such as theses, reports, workshop summaries, papers lacking experimental or analytical results, and non-English publications were excluded. After duplicate removal and excluding studies that used AIS only descriptively or where AIS played an incidental role, a total of 100 peer-reviewed journal articles and conference papers were retained for the final review.

Programming languages were not used as a screening criterion because the implementation language does not determine the scientific relevance, methodological soundness, or application domain of an AIS study. As with any literature review, the identified corpus may be influenced by database coverage and indexing practices, the choice of search terms, and language and publication bias. To mitigate these risks, searches were conducted across multiple sources, explicit inclusion/exclusion criteria were defined, and screening was performed in a structured and transparent manner. Nevertheless, relevant studies may remain underrepresented if they are not indexed in the searched sources or do not use standard AIS terminology in titles/abstracts. To complement the literature screening process, author keywords from the final set of studies were analyzed using VOSviewer (version 1.6.20). A customized thesaurus was applied to merge synonymous or closely related terms, and a minimum occurrence threshold of two was used to highlight salient research themes. The resulting co-occurrence network, Figure 2, demonstrates the most frequently associated and conceptually interconnected keywords, providing insight into the dominant topics and methodological directions within recent AIS-based research.

Table 1 summarizes the distribution of selected papers across journals and publishers, dominated by contributions from Elsevier, MDPI, and IEEE journals, with Ocean Engineering and the Journal of Marine Science and Engineering representing the most frequent publication venues. The geographical distribution of AIS-related research in Figure 3 shows that China has the highest publication output, contributing 23 studies independently and an additional 20 through international collaborations. Other countries with notable contributions and collaborations include South Korea, the United Kingdom, Norway, Portugal, Singapore, and Spain. This wide geographic spread highlights the global engagement in AIS-based research and reflects the shared importance of AIS data for advancing maritime safety, operational efficiency, and environmental sustainability across regions. Table 2 distinguishes analysis-oriented works (indicator-based, interpretable AIS analytics) from modeling-oriented works (predictive/optimization). The categories encompass collision risk assessment, anomaly detection, trajectory prediction, maritime traffic monitoring and management, ship behavior analysis, environmental impact assessment, wildlife conservation, maritime spatial planning, maritime surveillance and security, traffic complexity analysis, vessel classification, port operations, autonomous navigation, and general maritime analytics. Across applications, analysis approaches are typically more interpretable and data-light, whereas modeling approaches offer stronger predictive power but require robust validation and are more sensitive to domain shift and data quality.

2.2. General Workflow for AIS-Based Applications

The development of AIS-based applications typically follows a structured, multi-stage pipeline, summarized in Figure 4 and adapted from [76]. The first stage, data acquisition and preprocessing, focuses on collecting raw AIS messages and preparing them for reliable downstream use. AIS data are obtained primarily from two sources, terrestrial AIS (T-AIS) and satellite AIS (S-AIS), which differ substantially in coverage, reception frequency, message density, latency, and reliability. T-AIS typically provides denser, more continuous, and lower-latency observations in coastal waters, port areas, and nearshore traffic corridors, making it particularly suitable for fine-scale trajectory reconstruction, maneuver detection, port-call identification, and short-term behavioral analysis. In contrast, S-AIS offers much broader spatial coverage, especially in offshore and remote regions, but is more prone to sparse reception, message collision, irregular update intervals, and delayed availability. Because these source characteristics directly shape the structure and continuity of the observations, preprocessing is a necessary methodological stage rather than a purely technical preliminary step: data from both sources may be affected by missing transmissions, duplicated records, positional noise, message filtering, and irregular sampling intervals, so preprocessing commonly involves cleaning, filtering, interpolation, trajectory segmentation or reconstruction, and noise reduction before downstream analysis. These differences also imply that T-AIS and S-AIS should not be treated as interchangeable inputs, since the underlying acquisition mode influences the scale, reliability, and interpretability of subsequent modeling results.

The second stage, data analysis, examines the processed data to identify fundamental navigational, spatial, and temporal patterns. Typical approaches include clustering, density mapping, descriptive statistics, and exploratory spatiotemporal analysis, which help reveal movement behaviors, traffic flows, route structures, and operational characteristics of vessels in different maritime contexts. In the third stage, knowledge extraction [115], raw AIS trajectories are transformed into more informative and compact representations that are suitable for interpretation and modeling. This stage commonly includes trajectory compression, feature engineering, and the derivation of geometric, kinematic, and contextual descriptors, thereby converting raw vessel tracks into higher-level indicators of ship behavior and maritime activity. The fourth stage, modeling and inference, uses these extracted representations to build predictive, descriptive, or decision-oriented models. Classification, regression, probabilistic, and optimization-based approaches are employed to support a range of tasks, including anomaly detection, trajectory prediction, behavior recognition, and risk assessment [116,117]. At this stage, the objective is not only to characterize past and present vessel activity, but also to infer future states, detect abnormal situations, and support risk-aware maritime decision-making.

The final stage, runtime application and operational use, integrates these models into practical maritime systems and decision-support environments. In this stage, analytical outputs are translated into actionable services such as real-time alerts, traffic monitoring dashboards, collision-risk warnings, adaptive navigation assistance, and other operational tools that enhance maritime safety, efficiency, and governance (Figure 4: data processing → data analysis → knowledge extraction → modeling → runtime applications).

3. AIS Data Analysis Overview

AIS data analysis can be viewed in four perspectives: statistical analysis, safety indicators, situational awareness, and anomaly detection. These domains are inherently interconnected: statistical patterns inform the construction of safety indicators; safety metrics contribute to situational awareness; and deviations from expected behavior support anomaly detection. Together, they form the analytical foundation for understanding the dynamics of maritime traffic and supporting data-driven ship behavior modeling.

3.1. Statistical Analysis

Statistical analysis provides the quantitative base for interpreting large-scale vessel movement patterns using AIS data. Beyond descriptive measures such as speed distributions and port arrival times, modern AIS analytics incorporate trajectory clustering, density estimation, and probabilistic sequence models (e.g., Hidden Markov Model (HMM)) to infer common navigational routes and behavioral states. Traffic density maps and time charts expose areas of high traffic concentration, and spatiotemporal heatmaps reveal congestion hotspots and peak movement periods, while temporal analysis supports the estimation of port activity cycles and navigational bottlenecks.

These statistical outputs often serve as inputs to higher-level safety assessment and decision-support frameworks. In this context, the integration occurs at the level of maritime traffic monitoring, where AIS-derived statistical features are combined with navigational risk indicators to provide a more comprehensive interpretation of vessel behavior. The main elements involved in this integration include statistical outputs such as deviations from typical route corridors, anomalous speed variations, traffic density distributions, and clustering-based outliers, together with safety metrics such as the Distance to the Closest Point of Approach (DCPA) and Time to Closest Point of Approach (TCPA). In some cases, course over ground, speed over ground, heading, and positional updates from AIS messages are further used to derive encounter patterns and local risk conditions. By combining these elements, the analysis helps identify hazardous areas, abnormal ship interactions, and emerging conflict situations, thereby informing proactive traffic management, operational planning, and risk-based maritime decision-support systems. Thus, statistical analysis not only characterizes traffic flows but also contributes to integrated safety evaluation in maritime systems.

3.2. Safety Indicator

Safety indicators play a central role in quantifying the risk of vessel interactions and supporting collision avoidance. Fundamental geometric descriptors such as distance and relative bearing provide essential insight into vessel proximity and motion dynamics. In AIS-based maritime research, the Closest Point of Approach (CPA) concept is widely used to evaluate the potential interaction risk between vessels based on their reported motion states. Using AIS-derived information such as vessel position, speed over ground, course over ground, and heading, researchers can estimate relative trajectories and determine the predicted point of minimum separation between two vessels under the assumed motion. In this context, DCPA and TCPA are the standard quantitative indicators derived from CPA. DCPA estimates the minimum predicted distance between two vessels if they maintain their current motion states, while TCPA estimates the time remaining until that minimum distance is reached. Because these indicators can be directly computed from AIS kinematic data, they are frequently employed in AIS-based publications for collision risk assessment, encounter analysis, anomaly detection, and maritime traffic safety evaluation. Together, they provide interpretable and operationally relevant metrics for evaluating encounter severity.

Safety assessments extend beyond pairwise interactions to incorporate a vessel's maneuvering safety margin, the effective space required to execute safe maneuvers under given operational and environmental conditions. This margin depends on vessel maneuverability, hydrodynamic characteristics, traffic density, visibility, environmental forces, and meteorological conditions, as well as COLREGs obligations. For example, geometric and shape-based features extracted from AIS data have been applied to classify vessel behavior and detect anomalies [47]. Behavioral classification also supports safety monitoring. Xin et al. [79] distinguish between underway, anchorage, and at-berth states, while Sturgis et al. [83] identify drifting and anchored modes that may indicate elevated risk if occurring in constrained areas. These indicators collectively support early danger awareness, collision risk assessment, and compliance monitoring, forming a critical component of maritime safety systems.

3.3. Situation Awareness

Situation awareness (SA) refers to the perception, comprehension, and projection of elements in the maritime environment [118]. It underpins safe navigation, especially in dense and dynamic waterways. Figure 5 illustrates Endsley's three-level model adopted for AIS-based maritime traffic monitoring. At the perception level, AIS and radar provide real-time information on vessel identities, positions, speeds, and headings. The comprehension level integrates these elements to assess encounter situations, evaluate risk, and interpret vessel intent. At the projection level, future states such as predicted CPA/TCPA or potential rule violations are inferred to support decision-making (perception: detect ships via AIS/radar, observe speed and heading, monitor sea state and wind; comprehension: assess collision risk, interpret navigational intent, evaluate COLREGs traffic compliance; projection: forecast vessel behavior, anticipate congestion, plan avoidance maneuvers, project risk evolution).

Traffic state indicators such as static and dynamic density, encounter geometry, traffic complexity, and collision risk further enhance SA [64]. Encounter analysis is particularly relevant, as it classifies navigational interactions (head-on, crossing, and overtaking) in accordance with COLREGs, providing interpretable descriptors for vessel behavior and facilitating automated rule compliance assessment [119]. Graph-theoretic and network-based approaches increasingly support multi-ship encounter understanding, enabling recognition of collective movement patterns and emergent risks. Risk assessment models draw from geometric, probabilistic, and ML approaches. Bayesian networks capture dependencies among risk factors, ship domains and velocity obstacle models characterize spatial constraints, and ML models (e.g., RF, XGBoost, and NN) are widely used to classify risk levels and forecast collision likelihood. Hybrid methods that integrate deep learning with physical or probabilistic modeling have recently emerged as promising tools for improving prediction accuracy and robustness.

3.4. Anomaly Detection

Anomaly detection identifies deviations from expected vessel behavior and supports maritime surveillance, safety monitoring, and regulatory enforcement. Figure 6 illustrates the major categories of anomalies detectable from AIS data. Anomalies may be kinematic, such as irregular speed, unexpected course changes, or extreme turning rates. They can be positional, including off-route deviations or unexpected geographic locations. Contextual anomalies arise when a vessel's behavior is inconsistent with its type, operational norms, or environmental constraints. Data-related anomalies include AIS gaps, transmission manipulations, and spoofing. Liu et al. [120] further refine kinematic anomalies into speed/course anomalies (SCAs), turning anomalies (TAs), and loitering anomalies (LAs), enabling more structured analysis of vessel behavior. Applications span maritime safety, traffic management, security and enforcement, environmental monitoring, and real-time decision support.

Anomaly detection methods range from descriptive statistics and rule-based systems [121] to probabilistic sequence models [23,40] and ML approaches such as DBSCAN, K-means, GMMs, and supervised classifiers. Knowledge-based systems incorporate maritime regulations and expert rules, while DL methods, including LSTM networks, autoencoders, and CNN models, capture complex temporal dependencies and subtle deviations [122]. Recent advances include online stochastic modeling approaches such as the Ornstein–Uhlenbeck process for real-time detection [44]. A persistent challenge is balancing detection sensitivity with false alarm reduction, especially in regions with highly variable traffic patterns. Effective preprocessing, including noise filtering, trajectory segmentation, normalization, and map matching, significantly improves model performance and reduces misclassifications in dense maritime environments.

4. Modeling Taxonomy

Modeling ship behavior involves constructing mathematical or algorithmic representations that capture how ships move, interact, and respond to environmental and navigational conditions. In AIS-based research, modeling serves multiple purposes, including behavior recognition, trajectory prediction, anomaly detection, encounter classification, and decision support for collision avoidance and routing. Since AIS observations encode both kinematic information and interaction cues, ship behavior modeling has progressed toward increasingly data-driven, context-aware, and multi-agent frameworks. Ship behavior can be described from several perspectives, such as navigational status (e.g., underway, stopped, and anomalous), operational activities (e.g., docking, anchoring, and cargo handling), or interaction type (e.g., head-on, crossing, and overtaking) [78,117,123]. Additional taxonomies consider contextual factors (e.g., harbor vs. open sea), temporal variation, and behavioral scales ranging from individual vessels to multi-ship encounters. These perspectives highlight the need for a coherent modeling taxonomy capable of relating methodological choices to the specific characteristics of ship behavior problems.

To provide such a structure, this paper categorizes AIS-based modeling approaches into three foundational paradigms: classification, regression, and optimization. This taxonomy reflects the mathematical nature of behavior modeling tasks: assigning discrete behavior labels (classification), predicting continuous navigational states (regression), and computing optimal actions or routes under constraints (optimization). Table 3 compares these paradigms in terms of mathematical characteristics, strengths, limitations, and suitability for maritime applications. It is important to distinguish between modeling paradigms and algorithmic techniques. In this review, classification, regression, and optimization are treated as task-level paradigms, whereas methods such as DBSCAN, Random Forest, LSTM, GNNs, and DTW represent algorithmic techniques that may be incorporated within or across these paradigms depending on the target task. Figure 7 presents the distribution of modeling techniques used in recent AIS-based ship behavior studies: DBSCAN and deep learning methods appear among the most frequently used approaches, together accounting for nearly 40% of the models represented in the reviewed studies.

Classification: Classification models assign vessels, trajectory segments, or encounter situations to discrete behavioral categories based on AIS-derived features. In AIS-based studies, this paradigm is interpreted broadly to include not only supervised category assignment but also unsupervised grouping methods used for behavior pattern discovery, route clustering, and anomaly identification. These methods are used for ship-type identification, encounter classification, anomaly detection, risk-level assessment, and intent inference. Classification relies on decision boundaries, similarity relationships, or probabilistic formulations to distinguish among behavioral states. Representative techniques include SVM, Logistic Regression, DBSCAN, HMM, Decision Trees, RF [85], CNN-based classifiers [13,38], and multi-attention LSTM [119]. These models handle heterogeneous AIS features well, but face challenges such as class imbalance, ambiguous maneuver patterns, and hybrid behaviors not captured in training data. The prominence of DBSCAN suggests that a large part of the field still depends on unsupervised structure discovery, route extraction, and anomaly identification in settings where reliable behavioral labels remain scarce, while the parallel rise of deep learning reflects the increasing importance of nonlinear temporal prediction, high-dimensional sequence modeling, and context-sensitive inference. Traditional machine learning approaches, such as Random Forest, K-Means, and hybrid ensemble models, form a robust middle tier, indicating that interpretability, moderate data requirements, and computational efficiency remain important practical considerations. Emerging techniques such as Graph Neural Networks (GNNs), self-supervised learning, and Dynamic Time Warping (DTW)-based models suggest a gradual shift toward relational, representation-learning, and sequence-alignment perspectives, although these approaches have not yet reached the same level of methodological consolidation as clustering- and prediction-oriented models.

Regression: Regression methods learn continuous relationships from historical AIS time-series to forecast speed, position, course, ETA, or associated risk metrics. Techniques include RF regression, SVM regression, LSTM, GRU, ARIMA, Kalman filtering, and other sequence learning architectures. Regression is fundamental for trajectory prediction and proactive traffic monitoring but requires clean data, consistent sampling, and careful calibration to avoid error amplification. Representative examples include route prediction using combined DP-DTW clustering and regression frameworks [50] and ship prediction using RF models [109]. Optimization: Optimization approaches formulate navigation and decision-making tasks such as route planning, speed control, collision avoidance, and berth scheduling as constrained optimization problems [124]. These techniques include Reinforcement Learning, Convex Optimization, Genetic Algorithms, A∗ search, dynamic programming, and evolutionary strategies such as PSO-BP. Optimization is effective for generating COLREG-compliant trajectories and improving fuel efficiency, but often requires high-quality prediction inputs and can become computationally intensive, particularly in dense, multi-ship environments. Examples include quasi-intelligent route extraction from AIS data [92] and MPC-based path planning that integrates vessel dynamics and COLREGs for safe maneuvering [12]. Data-driven network inference from AIS trajectories has also been explored for large-scale digital shipping networks [62,91].

Taken together, these three paradigms address different analytical needs in AIS-based ship behavior research. Classification-oriented approaches are most suitable when the objective is to identify discrete behavioral states, operational categories, or anomalous events, but they are less effective for forecasting continuous vessel evolution. Regression-based models are more appropriate for trajectory prediction and risk estimation over time, although their performance is often sensitive to AIS sparsity, sampling irregularity, and abrupt maneuver changes. Optimization-based methods differ from the other two in that they are not primarily inferential, but prescriptive: they use predictions, constraints, and operational rules to generate actionable decisions such as route selection or collision-avoidance maneuvers. This distinction is important because it shows that AIS-based behavior modeling is not a single methodological problem, but a pipeline of related tasks spanning inference, prediction, and decision support. Overall, the proposed taxonomy highlights that AIS-based ship behavior modeling has evolved from an isolated method application toward a more differentiated analytical landscape in which discrete behavior recognition, continuous state prediction, and prescriptive navigation support are increasingly treated as distinct but interconnected problems. At the same time, the literature still reveals an important gap: although many studies report strong performance within individual tasks, fewer address how these paradigms can be integrated into coherent operational systems that remain robust under noisy data conditions, cross-region variability, and constrained-water maneuvering contexts.

5. Discussion

This section synthesizes the 2022–2024 AIS-based ship behavior literature. Emphasis is placed on (i) positioning relative to prior AIS surveys, (ii) trends, methodological shifts, and persistent gaps, and (iii) implications for runtime deployment and operational decision support, with particular attention to relevance for autonomous navigation and maneuvering in constrained waters. The contribution of this review lies in conceptual integration rather than simple expansion of the method inventory.

5.1. Workflow-Based Synthesis and Modeling Taxonomy

Several influential AIS surveys (e.g., Tu et al. [17], Xiao et al. [18]) established systematic reviews of AIS-enabled maritime analytics by organizing core tasks such as trajectory analysis, anomaly detection, and traffic pattern mining, while emphasizing AIS data characteristics and largely offline modeling pipelines. Building on these foundations, this review focuses on the 2022-to-2024 literature, highlighting recent methodological developments and the growing shift toward real-time analytics and operational decision support. Rather than classifying studies solely by task, this review is organized using a five-stage workflow (data processing, data analysis, knowledge extraction, modeling, and runtime applications), linking methodological choices to data conditions, modeling assumptions, and deployment feasibility. In addition, a modeling taxonomy is proposed to synthesize classification, regression, and optimization paradigms comparatively, clarifying their typical assumptions, outputs, and suitability for operational tasks under real-world constraints. This positioning helps clarify what has changed most recently and why particular modeling choices are emerging in current AIS-based ship behavior research.

This perspective also clarifies that no single methodological family is universally optimal. Clustering approaches are often effective for exploratory analysis and for settings with limited labeled data, but they depend strongly on feature design, similarity measures, and the degree to which behavioral patterns form separable groups. Deep learning approaches are advantageous when large, relatively consistent datasets are available and when complex temporal or nonlinear dependencies must be captured. However, they are typically more data-intensive, computationally demanding, and less interpretable. Probabilistic models are particularly useful when uncertainty and state transitions must be represented explicitly, but their performance may depend on simplifying assumptions about latent structure and transition dynamics. Method selection should therefore be guided not only by predictive performance, but also by data quality, sampling regularity, trajectory completeness, interpretability, and operational constraints. The workflow and taxonomy are also relevant from an application perspective. In port operations and constrained coastal environments, methods that perform well on dense nearshore AIS data and provide interpretable outputs are especially useful for congestion monitoring, berth allocation, port-call detection, and turnaround analysis. In autonomous navigation and maneuvering support, the emphasis shifts toward near-real-time processing, short-horizon trajectory prediction, and robust detection of rapid behavioral changes under uncertainty. At broader regional or global scales, such as maritime traffic monitoring, scalability and wide-area coverage become more important than fine-resolution maneuver detection, which increases the relevance of clustering, route extraction, and anomaly-screening approaches. This application-oriented view reinforces that methodological choices should be evaluated in relation to spatial scale, latency tolerance, uncertainty handling, and the characteristics of the available AIS data.

5.2. Trends, Methodological Shifts, and Persistent Gaps

Figure 2 shows AIS as the dominant hub in the keyword network, indicating that most studies continue to frame their contributions around AIS data itself rather than around narrower application labels. Three themes are particularly visible. The first is trajectory and pattern mining, reflected in terms such as trajectory mining, trajectory data mining, spatial clustering, spatial analysis, and maritime traffic network. This indicates continued interest in extracting structure from movement histories and traffic flows. The second is safety and monitoring, centered on anomaly detection and supported by relative motion and related traffic concepts, suggesting that detection-oriented analytics remain central to operational risk awareness. The third is autonomy-oriented research, where terms such as autonomous ship and intelligent navigation appear linked to the previous themes. This suggests a growing connection between AIS-based analytics and perception or decision pipelines for autonomous systems, although this area remains less central than surveillance and traffic-mining applications. Smaller terms such as cybersecurity, sentinel, and marine spatial planning point to diversification toward security and environmental or spatial-planning uses. Across the 2022–2024 literature, three methodological shifts stand out. First, there is a clear move toward machine-learning-based prediction and detection [42,44], including self-supervised and weakly supervised designs that reduce dependence on scarce labels [46]. Second, hybrid pipelines that combine rule-based maritime logic with learned models are increasingly used to balance interpretability, robustness, and predictive performance. Third, more studies now address deployment constraints directly, including streaming inputs, latency, and integration into operational workflows [76]. Compared with earlier reviews that emphasized AIS fundamentals and static modeling, recent work shows a stronger orientation toward real-time analytics and decision-support integration.

Despite this progress, several persistent limitations remain. AIS-based methods are inherently sensitive to signal loss, incomplete reception, irregular transmission intervals, positional noise, message filtering, and deliberate manipulation such as spoofing. These issues do not simply reduce data quality in a general sense. They can fragment trajectories, distort derived motion features, weaken anomaly detection, bias vessel profiling, and reduce sensitivity to rare or unusual behaviors. As a result, they directly affect downstream model outputs, uncertainty estimates, and the interpretation of inferred vessel activity. Generalizability remains another major challenge. Many methodological components are broadly transferable, including preprocessing routines, common trajectory features, general modeling frameworks, and evaluation principles. However, their performance and interpretation often depend strongly on regional conditions such as traffic density, vessel composition, coastal and port geometry, local navigational practices, regulatory context, and AIS coverage or data quality. Models, thresholds, and behavioral labels developed in one region may therefore not transfer directly to another without recalibration or revalidation. In this sense, transferability applies more readily to analytical frameworks than to model calibration or behavioral interpretation.

A further gap concerns evaluation and benchmarking. Comparability across studies remains limited because datasets, preprocessing choices, behavioral definitions, validation strategies, and reported metrics often differ substantially. Progress toward more cumulative and reproducible research will require shared or clearly documented reference datasets, more consistent evaluation protocols, and task-appropriate performance measures. In particular, future evaluation should better account for temporal structure, class imbalance, missing data, and cross-region transferability. Without such standardization, it remains difficult to determine whether reported performance gains reflect genuine methodological advances or differences in experimental design. Finally, transparency and explainability remain underdeveloped, especially in safety-critical applications where trust, accountability, and interpretability are essential for operational adoption [40]. This issue becomes even more important as AIS-based analytics move closer to real-time decision support and autonomous navigation.

5.3. Potential Runtime Applications of AIS-Based Models

Against this backdrop, three promising runtime application areas can be distinguished, each representing a different level of methodological maturity and operational readiness. The first concerns autonomous ships, where AIS is increasingly combined with onboard sensors such as radar, sonar, and LiDAR (Light Detection and Ranging) to support situational awareness, collision avoidance, and route planning. Models based on unsupervised anomaly detection [42], online anomaly detection and behavior classification [44], and multi-stage traffic complexity assessment for probabilistic collision detection [31] illustrate plausible pathways for integration into MASS navigation modules. The keyword structure centered on traffic, prediction, and risk suggests that AIS analytics are currently most mature for autonomy functions related to encounter assessment and tactical decision support. By contrast, maneuvering-specific behaviors at low speed and close range remain less visible in the current literature.

The second application area is remote monitoring and traffic management. These settings increasingly rely on AIS-driven analytics for real-time fleet awareness and broader maritime domain awareness. Studies such as [40,41,43,45,46] show that unsupervised and learning-based methods can support large-scale anomaly detection and vessel surveillance. Entropy-based and quantitative risk models [25,27,28] extend this potential toward traffic risk evaluation and safety planning. Relative to autonomy-oriented applications, this domain aligns more closely with the present strengths of AIS data, especially wide-area coverage and continuous track-based monitoring. Even so, deployment remains constrained by false positives, regional sensitivity, and dependence on the completeness and reliability of AIS transmissions. Operational usefulness, therefore, depends not only on algorithmic performance but also on robustness, calibration, and the ability to support human operators without overwhelming them with uncertain alerts.

The third application area concerns onboard decision-support systems. In this context, AIS-based models are combined with collision-risk prediction, trajectory forecasting, and contextual information to recommend maneuvers, improve routing efficiency, and support safer navigation decisions [21,47,125]. Recent work on anchorage risk assessment [26] and berthing safety evaluation [30] further indicates that AIS-based analytics are beginning to address operationally relevant scenarios in ports and harbor approaches. However, compared with remote surveillance systems, onboard decision support places greater demands on timeliness, interpretability, and user trust. A model may perform well in retrospective evaluation yet still be difficult to operationalize if its recommendations are not transparent, context-sensitive, or compatible with bridge-level decision processes. This highlights the gap between predictive capability and actionable usability, which remains insufficiently addressed in much of the current literature. Overall, the reviewed studies indicate a clear movement toward more operationally relevant AIS-based systems, driven by advances in machine learning, real-time analytics, and multi-sensor data fusion. However, runtime maturity differs substantially across application domains. Remote monitoring currently appears to be the most immediately deployable use case because it aligns well with the large-scale observational strengths of AIS data. By contrast, autonomous navigation and onboard decision support require a higher degree of contextual awareness, interpretability, and robustness under noisy real-world conditions. Wider deployment will therefore depend on progress in four areas: resilience to imperfect AIS data, cross-region generalizability, transparent model behavior, and integration with existing maritime decision-support infrastructures. Future work should focus not only on improving predictive accuracy but also on developing evaluation frameworks that better reflect operational complexity and real-world usability.

6. Conclusions

By consolidating recent advances in AIS-based ship behavior modeling, this review provides a structured synthesis of studies published in English between 2022 and 2024. Through a structured literature review process informed by PRISMA principles, relevant studies were identified and categorized, highlighting how statistical, visualization, and ML techniques are applied to improve maritime safety, situational awareness, and operational efficiency. From this body of work, a modeling taxonomy is derived that organizes classification, regression, and optimization approaches according to operational tasks such as vessel type identification, trajectory prediction, and route optimization.

The analysis revealed persistent challenges at both the data and modeling levels. AIS datasets remain affected by noise, missing values, inconsistencies, and uneven coverage, while modeling approaches face issues of scalability, interpretability, and parameter sensitivity. Although DL methods capture complex spatial–temporal patterns, their generalizability across regions and vessel types is constrained by limited access to large, high-quality, real-time or simulated datasets. These limitations underscore the need for robust, explainable ML frameworks and onboard support tools that can operate under real-time constraints with reliability and transparency.

This review highlights three emerging runtime application domains for AIS-based models. Autonomous vessels are increasingly integrating live AIS streams into navigation and collision-avoidance pipelines, enabling dynamic route adaptation. Shore-based monitoring platforms employ anomaly detection and behavior analysis to support maintenance planning, incident investigation, and regulatory compliance. Integrated decision-support systems combine AIS with environmental and operational data, such as weather and ocean forecasts, to optimize routing, scheduling, and fleet-level decision-making. These applications demonstrate the growing operational relevance of adaptive maritime analytics and their potential to translate theoretical modeling into operational practice.

Future research should advance AIS-based maritime intelligence by focusing on three interconnected priorities. First, integrating heterogeneous data sources, including AIS, weather, radar, and camera feeds, into unified architectures will enable robust safety indicators, early anomaly detection, and richer situational awareness. Second, explainable AI approaches tailored to maritime operations are essential to ensure transparency, auditability, and trust as autonomy expands. Third, progress will depend on multidisciplinary collaboration across engineering, data science, and human factors, supported by evolving regulatory frameworks and explicit consideration of human oversight in monitoring and validation. Together, these directions provide a pathway toward intelligent maritime systems that are resilient, trustworthy, and operationally aligned. A central finding of this review is that AIS-based ship behavior modeling is not converging toward a single dominant methodological solution. Instead, the field is evolving as a differentiated analytical landscape in which classification-, regression-, and optimization-based approaches address distinct but interconnected tasks of inference, prediction, and decision support. In conclusion, synthesizing current methodologies, characterizing emerging applications, and identifying open challenges, this paper provides a structured road map for the next generation of AIS-based ship behavior modeling. The insights presented here aim to guide researchers, system designers, and practitioners in developing maritime traffic systems that are efficient, transparent, and resilient to real-world demands.