Semantic feature extraction and classification inference of ship trajectories using coding sequences: A case study of the Taiwan strait

In the process of monitoring a myriad of vessels, maritime administrators recurrently confront situations where crucial static fields, such as vessel type, are frequently missing from automatic identification system (AIS) data. This omission presents a significant obstacle to effective safety management, prompting the need for a method to rectify incomplete data. In terrestrial traffic, constraints dictated by road widths do not apply to marine traffic, thus allowing vessels of an identical type to opt for routes that diverge by numerous nautical miles. This condition renders trajectory clustering insufficient. To tackle this issue, this paper proposes a division of the sea area into spatiotemporal grids and a transformation of vessel trajectories into a coded sequence of navigational grids. Word2vec embedding is merged with deep neural networks predicated on convolutional neural networks (CNNs), recurrent neural networks (RNNs), and attention mechanisms through the utilization of natural language processing techniques. The goal is to facilitate efficient vessel classification via deep learning. Applying the Taiwan Strait as a case study, it is shown that under the conditions of a five-dimensional word embedding, the F1 score for vessel classification using RNN models can attain 0.94. This investigation provides a unique method for processing vessel trajectory data, and it also assists in the effective completion of missing vessel types.