Multi-modal ship trajectory prediction via multi-scale temporal attention with diverse mode decoders

Accurate ship trajectory prediction supports encounter-risk screening and vessel traffic services, but latent intent makes future motion multi-modal across temporal scales. We propose Multi-Modal Multi-Scale Temporal Attention GRU (MM-MSTA-GRU), which adaptively fuses short, mid, and global temporal attention and uses K mode-conditioned decoders trained with Winner-Take-All and diversity objectives. On 26 days of Danish Maritime Authority AIS data ( ∼ 38,000 cargo and tanker trajectories), the deterministic variant achieves the best ADE, FDE, and MAE among the evaluated point predictors. The multi-modal variant yields 21.93 m minADE and 45.83 m minFDE over five minutes, increasing 100 m endpoint coverage from 87.0% ( K = 1 ) to 92.1%, while its highest-probability trajectory achieves 27.80 / 63.65 m ADE/FDE. On the geographically distinct NAIS harbor dataset, the predicted set obtains 73.33 m minFDE and 80.8% coverage. Mode-ranking probabilities are calibrated for oracle-best-mode selection (ECE 0.029), and the method outperforms four native multi-modal baselines in combined geometric and probabilistic quality. A conservative flag-if-any-mode rule recovers 72% of dangerous encounters missed by a point forecast in controlled tests and 70% on real held-out encounters—a recall-oriented operating point that trades precision for recall (higher false-alarm rate, lower F 1 ); the balanced probability-weighted set instead attains the best F 1 and PR-AUC at every safety radius at essentially unchanged decision accuracy. These results support a compact probabilistic predictor for maritime decision support.