Developing contextually aware ship domains using machine learning

Developing risk models to predict where collisions between vessels might occur is hindered by the relative sparsity of collisions. To address this, vessel encounters and near-misses have been used as a surrogate indicator of collision risk, referred to as ‘domain analysis’. When constructed empirically, using historical information, previous work is challenged by the multitude of factors which influence the passing distances between vessels. Within this paper, we conduct data mining of big vessel traffic datasets to determine the encounter characteristics across different waterways, vessel types and speeds, weather conditions and other exploratory variables. To achieve this, we utilise a novel approach of machine learning through a random forest algorithm to predict the critical passing distance between vessels in a multitude of conditions. We contribute a far greater range of influencing factors on domain size and shape than previous studies. Finally, we investigate the potential advantages of this approach to assess the spatial risk of collision across a large region. The results help to establish the factors that influence collision risk between navigating vessels and enable empirical maritime risk assessments.