Sailing the Seaformer: A Transformer-Based Model for Vessel Route Forecasting

The huge amount of traffic maritime data kicked off the challenge of their interpretations to predict the behavior of vessels during their trip. The availability of this type of data in large quantities is due to the Automatic Identification System (AIS), which is often used by ships for multiple reasons, such as national laws and security. We use this vast amount of AIS records to address the problem of vessel route forecasting, which is still tough to solve. In particular, we propose a novel deep learning architecture, the SeaFormer, which leverages the power of transformer modules to capture long-term dependencies, thus enabling the forecast even several hours ahead. The use of a Gumbel softmax allows to approximate the samples from a categorical distribution. Moreover, by leveraging the adopted activation function, we propose different sampling processes to enhance both the prediction of the vessel speed and position, with no change in the architecture of the model. The proposed method outperforms current state-of-the-art models in several scenarios using real data related to the Mediterranean Sea. The code is available at www.github.com/ispamm/SeaFormer.