Multi-source fusion for environment- and interaction-aware ship trajectory prediction

Ship trajectory prediction supports maritime safety, but AIS-based methods often neglect environmental constraints and vessel interactions. We propose Env-GATrans, which fuses target-vessel motion, a local navigation grid, and an ego-centered star interaction graph. AIS data from Danish waters are aligned with Copernicus Marine Service land-sea masks, bathymetry, and surface currents. A CNN encodes environmental features, a graph attention network models neighboring vessels, and a Transformer encoder-decoder predicts future positions. Across short-, medium-, and long-horizon tasks, Env-GATrans outperforms LSTM, Transformer, SNIINet, and an RNN Fusion + Surface Current baseline, reducing ADE and FDE by 7.80%–26.86%. Ablations show that interactions matter more at short horizons, while environmental features gain importance at longer horizons. Surface currents contribute most at short horizons; the land-sea mask becomes more influential over longer horizons, while bathymetry has a smaller marginal effect.