A Sequence-to-Sequence Based Muti-Semantic Network with Attention for Long-Term Vessel Trajectory Prediction

Forecasting ship trajectories is essential for ensuring the safety and efficiency of maritime operations. The evolution of deep learning has enabled using AIS (Automatic Identification System) data to predict vessel paths effectively. However, current approaches rely solely on a ship's latitude, longitude, and heading, leading to inaccuracies when forecasting diverse ship types. Additionally, a ship's sailing range is influenced by factors like endurance, fuel capacity, tides, and crew routines. To tackle these challenges, we introduce a new Sequence-to-Sequence based long-term vessel trajectory prediction network, named S2S-MSA. This model features a multi-semantic encoder-decoder architecture with an attention mechanism. The encoder translates voyage details into enriched latent vectors to capture navigational characteristics. We also introduce a multi-semantic decoder that considers ship type for trajectory forecasting. To assess S2S-MSA's performance, we use AIS data from ships in the southern Yangtze River estuary and compare it with other models. The outcomes show that S2S-MSA excels in both quantitative and qualitative metrics.