Data-Driven Prediction of Ship Destinations in the Harbor Area Using Deep Learning

Abstract

The prediction of ship destinations in the harbor can be utilized to identify future routes for navigating ships. The maritime traffic data are broadly classified into the ship trajectory data and the port information management data. These data have been accumulated for many years on the shore base station of different agencies, and are being utilized for evaluation of collision risk, prediction of vessel traffic , and other maritime statistical analysis. This paper presents a new destination prediction model of navigating ships in the harbor which consists of the candidate harbor proposal module and the position–direction filter module. The candidate harbor proposal module is trained by a deep neural network which makes use of the characteristics of ships and the occupancy distributions of piers. The position–direction filter module leaves out non-promising ones from the harbor list provided by the candidate proposal module, with respect to the current position and direction of navigating ship. In the experiments on real vessel traffic data, the proposed method has shown that its accuracy is higher than the frequency-based baseline method by about 10–15%.