AIS-Based Vessel Trajectory Prediction Using Memory... | AI Research

AIS-Based Vessel Trajectory Prediction Using Memory-Augmented Neural Networks
Accurate prediction of where a ship will be in the future is vital for maritime safety, collision avoidance, and efficient route planning. While researchers have successfully used "memory-augmented" neural networks to predict the paths of pedestrians and road vehicles, this technology has been largely untested for maritime traffic. This paper explores whether these memory-based systems—which mimic human behavior by recalling past experiences to inform future decisions—can improve the accuracy of vessel trajectory predictions using data from the Automatic Identification System (AIS).

How the Memory Approach Works

The researchers adapted a model called MANTRA, which uses an external memory bank to store examples of past vessel movements. The system works in three main stages: 1. Encoding: The model uses an autoencoder to turn trajectory data (position, speed, and course) into compact "encodings" that represent both past history and future outcomes. 2. Memory Construction: A controller evaluates these encodings. If a vessel’s movement is novel or difficult to predict, the system saves that specific pattern into an external memory bank. This allows the model to explicitly store rare maneuvering behaviors alongside common cruising patterns. 3. Retrieval: When predicting a new trajectory, the model compares the current vessel's path to the stored examples in its memory. It retrieves the most relevant past patterns to generate multiple possible future paths, accounting for the fact that vessels can often take different routes.

Incorporating Maritime Data

To make the model effective for ships, the authors modified the original approach to include specific maritime features. They integrated Speed Over Ground (SOG) and Course Over Ground (COG) as key inputs, as these are critical for understanding how a vessel moves. They also implemented a data preprocessing step to handle the common issue of missing AIS signals, using interpolation to fill gaps and a binary indicator to inform the model when data points were originally missing.

Performance Gains

The researchers tested their model using AIS data from the Gulf of Mexico and the New York Bight. They compared their memory-based approach against six established deep learning models, including those that use graph networks and Transformers. The results showed that the memory-augmented model consistently outperformed all other baselines. In the Gulf of Mexico, the model reduced the Average Displacement Error (ADE) by up to 46.4% and the Final Displacement Error (FDE) by up to 54.7% compared to the best-performing alternative.

Key Takeaways

The study demonstrates that memory-based prediction is a highly effective, stable, and accurate paradigm for maritime navigation. Notably, the model achieved these superior results without explicitly modeling interactions between different vessels, suggesting that the ability to "remember" and retrieve specific past movement patterns is a powerful tool for predicting future vessel behavior.