An adaptive vessel trajectory prediction network with feature-time two-dimensional attention and learnable complex spectral modulation

Multi-step vessel trajectory prediction remains challenging because critical historical observations are unevenly informative, local nonstationary maneuvers are difficult to characterize, and temporal and spectral representations are often insufficiently coordinated. To address these issues, this study proposes 2D-LSMGNet, where 2D, LSM, and G denote feature-time two-dimensional attention, learnable complex spectral modulation, and adaptive gating, respectively. A two-layer bidirectional long short-term memory (Bi-LSTM) network with two-dimensional attention learns separate temporal dependencies for latent feature channels. In parallel, short-time Fourier transform and learnable complex spectral modulation jointly adjust the amplitudes and phases of localized spectral components. Adaptive gating coordinates both representations at each historical time step and latent channel, followed by future step queries, cross-attention, and an LSTM decoder. Using 3 h of Danish AIS observations to predict the subsequent 1 h, 2D-LSMGNet achieves an ADE of 0.7695 nmi and an FDE of 1.7539 nmi, reducing them by 13.6% and 13.3% relative to the strongest baseline. Direct evaluation on unseen 2025 Guam AIS data yields 0.9737 nmi ADE and 2.1744 nmi FDE without parameter updating. Ablation and interpretability analyses further demonstrate the contributions of joint amplitude and phase modulation and adaptive temporal–spectral coordination.