InlandVessel-2K-Sim: A Reproducible Multi-Source Testbed and Evaluation Protocol for Inland Vessel Trajectory Prediction, with a Spatiotemporal Graph Attention Reference Model

Abstract We release InlandVessel-2K-Sim , a configurable simulator generating synchronised AIS-radar-vision scenes for two inland channel geometries in the field-recording data layout, with an evaluation protocol anchoring absolute error on constant-velocity extrapolation, averaging three seeds on a fixed split, and scene-level bootstrap and paired tests. The testbed is demonstrated with ST-GAT-Fusion , a spatiotemporal graph attention reference model coupling three modality-specific encoders through dynamic-time-warping alignment, with explicit vessel-to-vessel and vessel-to-shore edges. On the released 160-scene split the four fusion-and-graph configurations sit at 19.2–19.5 m ADE — roughly half the 40.20 m of constant-velocity extrapolation — with differences inside seed variability (0.62–0.89 m for fused models, 0.04–0.11 m for AIS-only baselines). We therefore report, as a primary finding, that multi-source fusion yields no resolvable accuracy advantage: a capacity-matched AIS-only Transformer matches the fused model, a fixed-budget data-scale control shows the fusion edge eroding rather than growing with training scenes, and a larger AIS-only Transformer reaches 17.3 m, below all fused configurations. What it delivers is robustness: under AIS outage the fused model loses less than half a metre where the AIS-only baseline loses ten, and with no AIS it still predicts while the baseline cannot. The half-width model sustains 179 vessels/s on CPU.