Digital maritime monitoring: Enhancing situational awareness in shipping traffic using AI-based models
Farshad, Farahnakian (2026-02-27)
Digital maritime monitoring: Enhancing situational awareness in shipping traffic using AI-based models
(27.02.2026)
Turun yliopisto
Julkaisun pysyvä osoite on:
https://urn.fi/URN:ISBN:978-952-02-0527-0
https://urn.fi/URN:ISBN:978-952-02-0527-0
Kuvaus
ei tietoa saavutettavuudesta
Tiivistelmä
Maritime Situational Awareness (MSA) is crucial for the safety and security of maritime operations. It depends on advanced sensing technologies, such as the Automatic Identification System (AIS), which enables continuous tracking data. However, the large volume of AIS messages transmitted by numerous vessels poses challenges for traditional analysis methods, which are often inefficient and unable to provide real-time monitoring and accurate insights. To address these challenges, it is necessary to integrate digital services equipped with AI-based predictive models and data-driven solutions to automate maritime monitoring tasks.
This thesis presents the development and evaluation of two key AI-driven services: (1) Ship Abnormal Behavior Detection (ShABD) and (2) Ship Movement Prediction (ShMP). The ShABD service targets a range of crucial maritime scenarios, including detecting dark ships, identifying unexpected changes in movement, recognizing spiral maneuvers, and uncovering potential smuggling activities. Since these behaviors threaten maritime security, effective, data-driven approaches are critical for timely detection and response. To address these threats, this work uses clustering-based algorithms, thereby eliminating the need to label millions of AIS observations manually. For the ShMP service, a variety of advanced Machine Learning (ML) approaches are investigated, including sequence-to-sequence models for short-term forecasting and similarity-based measures for predicting long-term trajectories.
The design of the ShMP frameworks enables two types of predictions, each catering to different stakeholder requirements. For example, port authorities benefit from comprehensive forecasts of ship movements to streamline loading and unloading operations, while coast guards require accurate short-term predictions to support timely and effective interventions.
The research presented in this thesis utilizes historical AIS data from ships operating in the Baltic Sea between 2022 and 2024. For data, two reliable sources were used: (1) HELCOM and (2) Digitraffic APIs. The main findings can be summarized as follows: K-Means outperformed DBSCAN, Gaussian Mixtures, and Affinity Propagation in detecting dark ships and spiral maneuvers, with an average silhouette coefficient of 0.755. The use of 3D input features (latitude, longitude, speed, or course) enhanced anomaly detection by more effectively separating normal and anomalous vessel behaviors. Furthermore, the proposed ship-trajectory forecasting models demonstrated strong performance. The short-term ShMP model, which uses a feed-forward neural network, achieved mean absolute errors (MAEs) of 0.05-0.13 for three-hour predictions and accuracies of 82–99% across different ship types. For long-term forecasts, the ShMP model utilized the Symmetrized Segment-Path Distance (SSPD) to find historically similar routes, achieving similarity scores of 0.1 and 0.11 for 10-hour predictions.
Moreover, integrating the long-term ShMP model with the AI-ARC maritime surveillance system improved smuggling detection rates, particularly in complex scenarios such as ship-to-ship transfers and unexpected route deviations. Although reliable smuggling detection cannot rely solely on AIS positional data, distinct features suggesting smuggling activity have been observed. In addition, a new version of the ShMP service was successfully evaluated with optimized look-back window sizes for 1-hour and 5-hour forecasting intervals in the Baltic Sea, demonstrating an improved ability to select key AIS messages for accurate ship movement prediction. The service leverages a look-back window size determination algorithm to train a Deep Learning (DL) model—specifically, the Temporal Convolutional Network (TCN)—thus optimizing the computational resources required for sustained operational efficiency.
The AI-based services developed in this thesis were further validated during the AI-ARC project demonstration event, utilizing real-time AIS data from ships operating in the Baltic Sea, particularly near Malm¨o, Karlskrona, and Copenhagen. Both the ShMP and ShABD services are poised to deliver significant benefits to maritime operations by providing actionable insights and advanced predictive analytics for maritime traffic management.
This thesis presents the development and evaluation of two key AI-driven services: (1) Ship Abnormal Behavior Detection (ShABD) and (2) Ship Movement Prediction (ShMP). The ShABD service targets a range of crucial maritime scenarios, including detecting dark ships, identifying unexpected changes in movement, recognizing spiral maneuvers, and uncovering potential smuggling activities. Since these behaviors threaten maritime security, effective, data-driven approaches are critical for timely detection and response. To address these threats, this work uses clustering-based algorithms, thereby eliminating the need to label millions of AIS observations manually. For the ShMP service, a variety of advanced Machine Learning (ML) approaches are investigated, including sequence-to-sequence models for short-term forecasting and similarity-based measures for predicting long-term trajectories.
The design of the ShMP frameworks enables two types of predictions, each catering to different stakeholder requirements. For example, port authorities benefit from comprehensive forecasts of ship movements to streamline loading and unloading operations, while coast guards require accurate short-term predictions to support timely and effective interventions.
The research presented in this thesis utilizes historical AIS data from ships operating in the Baltic Sea between 2022 and 2024. For data, two reliable sources were used: (1) HELCOM and (2) Digitraffic APIs. The main findings can be summarized as follows: K-Means outperformed DBSCAN, Gaussian Mixtures, and Affinity Propagation in detecting dark ships and spiral maneuvers, with an average silhouette coefficient of 0.755. The use of 3D input features (latitude, longitude, speed, or course) enhanced anomaly detection by more effectively separating normal and anomalous vessel behaviors. Furthermore, the proposed ship-trajectory forecasting models demonstrated strong performance. The short-term ShMP model, which uses a feed-forward neural network, achieved mean absolute errors (MAEs) of 0.05-0.13 for three-hour predictions and accuracies of 82–99% across different ship types. For long-term forecasts, the ShMP model utilized the Symmetrized Segment-Path Distance (SSPD) to find historically similar routes, achieving similarity scores of 0.1 and 0.11 for 10-hour predictions.
Moreover, integrating the long-term ShMP model with the AI-ARC maritime surveillance system improved smuggling detection rates, particularly in complex scenarios such as ship-to-ship transfers and unexpected route deviations. Although reliable smuggling detection cannot rely solely on AIS positional data, distinct features suggesting smuggling activity have been observed. In addition, a new version of the ShMP service was successfully evaluated with optimized look-back window sizes for 1-hour and 5-hour forecasting intervals in the Baltic Sea, demonstrating an improved ability to select key AIS messages for accurate ship movement prediction. The service leverages a look-back window size determination algorithm to train a Deep Learning (DL) model—specifically, the Temporal Convolutional Network (TCN)—thus optimizing the computational resources required for sustained operational efficiency.
The AI-based services developed in this thesis were further validated during the AI-ARC project demonstration event, utilizing real-time AIS data from ships operating in the Baltic Sea, particularly near Malm¨o, Karlskrona, and Copenhagen. Both the ShMP and ShABD services are poised to deliver significant benefits to maritime operations by providing actionable insights and advanced predictive analytics for maritime traffic management.
Kokoelmat
- Väitöskirjat [3086]