Spatial Modeling of Maritime Risk Using Machine Learning

Abstract Managing navigational safety is a key responsibility of coastal states. Predicting and measuring these risks has a high complexity due to their infrequent occurrence, multitude of causes, and large study areas. As a result, maritime risk models are generally limited in scale to small regions, generalized across diverse environments, or rely on the use of expert judgement. Therefore, such an approach has limited scalability and may incorrectly characterize the risk. Within this article a novel method for undertaking spatial modeling of maritime risk is proposed through machine learning. This enables navigational safety to be characterized while leveraging the significant volumes of relevant data available. The method comprises two key components: aggregation of historical accident data, vessel traffic, and other exploratory features into a spatial grid; and the implementation of several classification algorithms that predicts annual accident occurrence for various vessel types. This approach is applied to characterize the risk of collisions and groundings in the United Kingdom. The results vary between hazard types and vessel types but show remarkable capability at characterizing maritime risk, with accuracies and area under curve scores in excess of 90% in most implementations. Furthermore, the ensemble tree‐based algorithms of XGBoost and Random Forest consistently outperformed other machine learning algorithms that were tested. The resultant potential risk maps provide decisionmakers with actionable intelligence in order to target risk mitigation measures in regions with the greatest requirement.