Ship classification based on trajectory data with machine-learning methods

Abstract

Determining the type of a vessel solely by trajectory data is a desirable capability with many potential applications, however it is also a nontrivial task. In this paper, various machine-learning techniques are combined to train a model which is able to achieve this goal. In order to acquire training data, Automatic Identification System (AIS) messages collected from terrestrial and satellite base stations have been converted into ship trajectories including corresponding ship types. Since AIS is error-prone, preprocessing is applied to prepare the trajectories and remove errors from the dataset. Subsequently, we introduce a new set of features which contains behavioural and geographical properties, as well as daytime context information. Based on the generated features, a classification algorithm is trained to distinguish between five types of vessels: Cargo, Tanker, Passenger, Fishing and Other. Additionally, the influence of vessel dimensions as discriminative features is analyzed.