Maritime Behaviour Anomaly Detection with Seasonal Context

Abstract

Monitoring maritime traffic has become an important task for ensuring the safety of vessels, as well as the goods, and persons that they may be transporting. An active area of research is the modelling of expected normal vessel behaviour so as to detect subsequent anomalies in new data. Anomalies indicate that a vessel is not behaving in an expected manner and their detection can be flagged for further investigation to identify whether the vessel needs assistance or intervention. An important factor for some vessels in determining normal behaviour is seasonal context. However, current approaches typically do not incorporate seasonality into the model.In this paper, an approach is presented where seasonal context is incorporated into the behaviour model. Seasonal context is first incorporated into vessel trajectory data by encoding the month of year into a historic dataset. Following this, a model of normal behaviour is generated using a clustering approach, with DBSCAN used in this paper. Details of setting the DBSCAN parameters appropriately for vessel trajectory data are provided and four distance metrics explored. Resulting cluster models are evaluated in the context of using the model to classify previously unseen data as either fitting the model or constituting an anomaly. The experiments focus on using fishing vessels within two identified seasons to build the normal model, which is evaluated with a mixture of in season and out of season fishing and non-fishing vessel behaviour.