Author Identifier

Travis Rybicki: http://orcid.org/0000-0002-5661-2757

Date of Award

2025

Document Type

Thesis

Publisher

Edith Cowan University

Degree Name

Master of Computing and Security by Research

School

School of Science

First Supervisor

Martin Masek

Second Supervisor

Peng Lam

Abstract

Monitoring maritime traffic is essential for ensuring the safety of vessels, safeguarding transported goods or persons, and preventing illicit or hazardous activity at sea. Increasingly, researchers have explored data-driven approaches to model expected vessel behaviour and detect deviations or anomalies. These anomalies—such as course deviations, unauthorised area entries, or unexpected operational patterns—can indicate emergencies, regulatory breaches, or unlawful intent. Data broadcast by vessels provides a valuable resource for such analyses; however, the inherent complexity and context-dependency of maritime behaviour present persistent modelling challenges. One critical yet underutilised factor in this context is seasonality. For certain vessel types, for example, fishing vessels, patterns of activity vary predictably with the time of year. Ignoring these seasonal trends can lead to models that misclassify normal seasonal behaviour as anomalous, or worse, fail to detect meaningful outliers.

This research study proposes a two-phase approach for maritime anomaly detection while incorporating seasonal context. In the first phase, clustering techniques are applied to vessel trajectory data to identify patterns of normal operation. In the second phase, a classification process evaluates new trajectories against these models to detect anomalous activity. Central to this approach is the inclusion of seasonal context: month-of-year (MoY) values were encoded into each trajectory record, enabling the model to be seasonally aware.

A persistent challenge in trajectory similarity calculation lies in the variable length of trajectories and the appropriate choice of distance measures. To address this, four trajectory distance measures were assessed: Dynamic Time Warping (DTW), two normalised DTW variants, and the Hausdorff distance. Particular attention was given to DTW’s known sensitivity to trajectory length. The use of normalised variants was investigated to mitigate length-related biases and improve comparability between trajectories of differing durations.

The core seasonal model was constructed using a set of in-season-fishing vessel trajectories representing typical operational behaviour during active fishing periods. DBSCAN, a density-based clustering algorithm, was employed to identify coherent groups within this data, with parameter tuning based on data-driven distance thresholds. The model’s performance was then evaluated against a mixed set of trajectories comprising in-season and out-of-season fishing vessels, alongside cargo, passenger, and tanker vessels, to assess its capacity to detect both seasonal deviations and category-based anomalies.

Following clustering, a k-nearest neighbours (kNN) classifier was applied to classify new, unseen trajectories relative to the established seasonal models. Three distance-based thresholding metrics were tested: Cluster-Mean, Cluster-IQR-Rule, and DBSCAN Epsilon. Experimental results indicated that the Hausdorff distance and both normalised DTW variants consistently outperformed the non-normalised DTW measure, which remained unstable due to trajectory length sensitivity.

Finally, an ensemble framework, Inter-Model Reasoning (IMR), was introduced to combine the individual classifiers while excluding unstable distance measures. The IMR configuration delivered improved detection consistency, reduced false positives, and enhanced decision robustness at moderate consensus thresholds. These findings reinforce the value of incorporating seasonal context, adaptive distance measures through normalisation, and ensemble-based reasoning within maritime anomaly detection frameworks.

DOI

10.25958/rwqh-j475

Download

Share

 
COinS