ITRANSFORMER-BASED AIS TRAJECTORY PREDICTION MODEL WITH SPATIAL-TEMPORAL EMBEDDING

Improving the accuracy of automatic identify system (AIS) trajectory prediction is crucial for comprehensive maritime surveillance. Previous prediction model such as recurrent neural network (RNN), spatio-temporal graph convolutional networks and Transformer family models struggle to effectively handle long-term trajectory prediciton. Inspired by the latest advancements in long-term series forecasting (LTSF), we develop an improved iTransformer model based on spatial-temporal embedding. We conduct a 30 min trajectory prediction experiment on the AIS dataset provided by the Danish Maritime Authority, the proposed model achieved inspiring results. Additionally, we demonstrate the model’s potential for maritime surveillance by integrating AIS trajectory prediction with synthetic aperture radar (SAR) images.