ST-FEiTNet: A trajectory prediction model for complex sea states via multiscale spatiotemporal feature extraction and cross-modal frequency enhancement

Vessel trajectory prediction plays a pivotal role in maritime safety enhancement and energy-efficient navigation. However, existing methods struggle to achieve robust spatio-temporal modeling under complex marine environments due to insufficient integration of heterogeneous environmental factors. To address this limitation, this paper proposes a Spatio-Temporal Frequency-Enhanced iTransformer Network (ST-FEiTNet) for intelligent trajectory prediction under various sea states. First, we design a multi-scale spatio-temporal feature extraction module (MST-CTRG) to extract to jointly capture multi-scale temporal patterns and spatial social interaction feature. Second, we perform FECAM to model inter-feature dependence and dynamically integrate multi-modal features. Finally, we modify iTransformer by adding MST-CTRG and FECAM to predict vessel trajectory. Additionally, we propose a “motion-environment” joint representation to integrating the motion state (SOG, COG) with the dynamic characteristics of the sea state (wind, wave, current, and omaod550nm), we construct multi-dimensional features to enhance the simulation of the vessel's dynamic behavior in real ocean scenarios. Based on extensive experiments, ST-FEiTNet improves performance by 12.8 %, 15.2 % and 16.6 % in ADE over iTransformer for 1-min, 5-min and 10-min prediction. Introduction of “motion-environment” joint representation improved by 4.9 %, 7.1 % and 8.6 % in ADE for the same prediction horizons.