Long-horizon vessel trajectory prediction for maritime traffic monitoring and route planning requires both a prompt response to manoeuvres and control over error accumulation. To address these demands, researchers evaluated UnifiedInternalMoE, a framework combining a motion-prior residual expert with four temporal predictors in a separate checkpoint for each geographic region and seed.
The system fits convex weights to validation predictions at five, ten, and sixty minutes, subsequently interpolating these values to generate a complete future trajectory. Investigators evaluated the heterogeneous combination through a ten-seed comparison across four regions alongside capacity controls, learned gating, and internal ablation tests.
In a five-seed evaluation on an independent Finnish automatic identification system subset, the fusion model demonstrated the lowest mean average displacement error across all six evaluated horizons. At the sixty-minute threshold, the system achieved an average displacement error of 1719.53±25.80 meters.
Despite this accuracy, the architecture did not consistently lead in long-horizon final displacement error or in frozen source-to-target transfer evaluations. The findings support retraining the identical architecture across different regions as an analytical forecasting reference for shore-based tracking of long-horizon motion trends.
The authors noted clear operational constraints on the system’s broader deployment. Its real-world applicability remains limited by regional distribution differences, the total number of independent vessels, and substantial computational costs.
Subject of Research: Maritime vessel trajectory prediction using a heterogeneous mixture-of-experts architecture
Article Title: Heterogeneous mixture-of-experts for long-horizon vessel trajectory prediction across regional datasets
Article References: Tan, Y., Cai, Y., Fan, K., Li, X., & Qiu, G. (2026). Heterogeneous mixture-of-experts for long-horizon vessel trajectory prediction across regional datasets. PLOS One, 21(11), e0363079. https://doi.org/10.1371/journal.pone.0363079
Image Credits: AI Generated
DOI: 10.1371/journal.pone.0363079
Keywords: vessel trajectory prediction, maritime traffic monitoring, route planning, mixture of experts, motion-prior residual expert, temporal predictors, average displacement error, final displacement error, Finnish AIS dataset, trajectory forecasting, shore-based tracking, convex weighting, trajectory interpolation, motion trends, computational cost, regional adaptation
Cite Scienmag News
Gavin Prescott. (November 20, 2026). Vessel trajectory model cuts displacement error across regional navigation forecasts. Scienmag. https://scienmag.com/vessel-trajectory-model-cuts-displacement-error-across-regional-navigation-forecasts/
Gavin Prescott. "Vessel trajectory model cuts displacement error across regional navigation forecasts." Scienmag, 20 November 2026, https://scienmag.com/vessel-trajectory-model-cuts-displacement-error-across-regional-navigation-forecasts/. Accessed 20 November 2026.
Gavin Prescott. "Vessel trajectory model cuts displacement error across regional navigation forecasts." Scienmag. November 20, 2026. https://scienmag.com/vessel-trajectory-model-cuts-displacement-error-across-regional-navigation-forecasts/








