Risky maritime encounter prediction via ensemble machine learning

Abstract

As the demand for maritime transport is enhancing, ship-ship encounters are increasing. Narrow and congested waterways are primarily risky areas for potential consequences. The purpose of this study is to predict non-accidental risky encounters between two ships without distance between ships as a model variable on narrow and congested waterways via machine learning. A novel framework is developed to model risky encounters as predictable events. Site specific ship domain approach is used as the risk labeling criteria. To overcome imbalance in the nature of non-accidental critical events, ensemble machine learning algorithms are adopted. Historical navigational statistics of different sectors of the waterway are integrated to prediction model. The approach is tested on a historical AIS dataset from the Strait of Istanbul. To evaluate the methodology, accuracy, precision, recall and roc-auc metrics are used. K-fold cross validation and permutation based feature importance tests are performed. Each 4 out of 5 risky encounters are successfully predicted. Developed methodology can be used by vessels as an early potential collision avoidance alert system.