Maritime trajectory prediction using Bi-LSTM and Bi-GRU models

Abstract

Predicting vessel trajectories is essential for identifying illegal, unreported, and unregulated (IUU) fishing activities. Such identification requires Automatic Identification System (AIS) data to be accurate, complete, and reliable. This work presents a maritime trajectory prediction system based on AIS data, using two bidirectional recurrent neural networks: Bi-LSTM and Bi-GRU, implemented in both Sequence-to-One and Multi-Step variants. The data was preprocessed through outlier removal, normalization, and standardization to ensure spatial consistency. The models were integrated into a Mercury-based web interface to allow easy visualization and analysis of the predicted trajectories. Evaluation using a custom geodesic loss function and Hausdorff distance shows that Bi-LSTM excels in modeling long-term dependencies on large datasets, while Bi-GRU provides competitive accuracy with superior computational efficiency. The choice between Sequence-to-One and Multi-Step architectures depends on vessel behavior and prediction context, highlighting that model selection should be task-driven.