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

Every day, thousands of cargo ships navigate the narrow, winding waterways of China's inland rivers, moving goods along vital shipping routes. As traffic grows denser, the need to predict where these vessels will go in the next few seconds becomes critical for preventing collisions, managing locks, and ensuring safety. For decades, the standard way to track these ships has been the Automatic Identification System, or AIS, a digital broadcast that tells a central computer a ship's location and speed. However, this system has a fatal flaw: the signal often disappears when a ship passes under a bridge, enters a lock, or sails near a crowded shoreline. Furthermore, small boats like fishing vessels often lack the equipment to broadcast a signal at all. When the signal vanishes, the computer is left guessing, and in a crowded channel, a guess can be dangerous.

To solve this, researchers have long hoped that combining the AIS signal with other senses—like radar, which sees through fog, and cameras, which see the shape of the boat—would create a perfect, unbreakable view of the water. The idea is that if one sensor fails, the others can pick up the slack, creating a safety net that is far more accurate than any single source alone. This concept, known as multi-source fusion, has become a standard approach in fields like autonomous driving, where cars combine cameras and radar to see the road. But does this same strategy work for the unique, chaotic environment of inland waterways, where ships are packed tightly together and move in complex, interdependent patterns?

A team of researchers from Jining Polytechnic and Qufu Normal University set out to answer this question with a new approach. Instead of relying on messy, real-world data that is hard to compare, they built a sophisticated digital simulator called InlandVessel-2K-Sim. This tool creates thousands of realistic scenes of inland waterways, complete with two different channel shapes, bridges, and fog. Crucially, it generates synchronized data from three sources at once: the AIS broadcast, radar scans, and camera vision. This allowed the team to run controlled experiments where they could test exactly how well a computer model performs when it uses all three senses versus when it relies on just the AIS signal. They also developed a strict set of rules for testing these models, ensuring that every experiment was measured against the same baseline: a simple guess that assumes the ship will continue moving at its current speed.

The researchers trained a complex artificial intelligence model, which they named ST-GAT-Fusion, to predict ship paths using this simulated data. This model was designed to be a "super-observer," capable of fusing the AIS, radar, and vision streams into a single, coherent picture of the waterway. They expected that by combining these three sources, the model would significantly outperform a system that only looked at the AIS signal. The results, however, were surprising. When tested on a fixed set of 160 simulated scenes, the multi-source fusion model did not predict ship paths any more accurately than a simpler model that used only the AIS signal. In fact, the most accurate model in the entire study was a large, powerful AI that used only the AIS signal, achieving a prediction error of about 17.3 meters over a 12-second horizon. The fusion model, despite its complexity and access to radar and vision, hovered around 19.2 to 19.5 meters of error. The difference between them was so small that it fell within the range of normal computer variation, meaning the extra sensors provided no measurable boost in accuracy for this specific task.

The study did not find that the extra sensors were useless, but rather that their value depends entirely on the situation. The researchers discovered that the fusion model's true strength was not in being more precise when everything was working, but in being more reliable when things went wrong. When the researchers simulated a scenario where the AIS signal was completely blocked—mimicking a ship passing under a bridge—the AIS-only model failed entirely because it had no data to work with. The fusion model, however, continued to predict the ship's path using only the radar and camera data, losing less than half a meter of accuracy compared to its normal performance. This robustness is a vital safety feature, proving that while the extra sensors may not make the system smarter in perfect conditions, they keep it from going blind when the primary signal fails.

The team also investigated why the fusion model did not become more accurate as they added more training data. They found that the advantage of using multiple sensors did not grow with more information; instead, the gap between the fusion model and the simple AIS-only model actually shrank as the training data increased. This suggests that the current methods for combining these different types of data are not yet efficient enough to extract extra value from the radar and vision streams in this specific environment. The study concludes that for the current size of the dataset, the complexity of fusing three sensors does not yield a better prediction than a well-tuned system using just the AIS signal. The real breakthrough of this work is not a new, super-accurate predictor, but the creation of a reproducible testing ground and a clear set of rules that revealed this truth. By providing a shared simulator and a rigorous evaluation protocol, the researchers have given the scientific community a way to test future ideas fairly, ensuring that new technologies are judged on their actual ability to improve safety rather than on the promise of combining more data.

In the end, the research offers a nuanced lesson for the future of maritime safety. While the dream of a perfect, multi-sensor vision system remains appealing, the reality is that simply adding more data does not automatically make a system smarter. The most effective path forward may involve focusing on making the existing AIS-based systems more robust and reliable, while using radar and vision specifically as a backup to prevent total failure when the primary signal is lost. The study establishes that in the crowded, complex world of inland waterways, the value of a new technology is not just in how much data it can process, but in how well it handles the moments when that data disappears.