🤖 AI

AIS-Based Vessel Trajectory Prediction Using Memory-Augmented Neural Networks

This paper demonstrates that memory-augmented neural networks significantly outperform standard deep learning baselines in predicting vessel trajectories using AIS data from the Gulf of Mexico and New York Bight, highlighting the value of external memory mechanisms for maritime safety and efficiency.

Original authors: Wonmo Koo, Sanha Chang, Heeyoung Kim

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wonmo Koo, Sanha Chang, Heeyoung Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Teaching Computers to "Remember" Like Humans

Imagine you are trying to guess where a friend will walk next. If you've seen them walk this route a hundred times before, you know they usually turn left at the coffee shop. But if you've never seen them turn right, you might be surprised when they do.

Humans are great at this because we have memory. We don't just calculate math; we recall past experiences to guess the future.

This paper asks a simple question: Can we give computers a similar "memory" to predict where ships will go?

The researchers found that yes, giving a computer an external "notebook" to store past ship journeys makes it much better at predicting where ships will be next, even better than the current top-tier computer models.

The Problem: Predicting Ship Paths

Ships move around the ocean, and we track them using a system called AIS (Automatic Identification System). It's like a digital logbook that constantly updates a ship's location, speed, and direction.

Predicting where a ship will go is tricky because:

Most of the time, ships just cruise. They go straight at a steady speed (like a commuter driving to work).
Sometimes, they do weird things. They might turn sharply to avoid a storm or another ship (like a driver suddenly swerving to avoid a pothole).

Current computer models are good at the "cruising" part but often get confused by the "weird" maneuvers. They try to learn the rules of the road, but they forget the specific details of past trips.

The Solution: The "Memory-Augmented" Computer

The researchers used a special type of AI called a Memory-Augmented Neural Network (MANN).

Think of this AI as a student taking a test:

Standard AI: This student tries to solve the problem using only the rules they memorized in class. If the question is tricky or unusual, they might guess wrong.
The New AI (MANTRA): This student has a cheat sheet (an external memory) filled with examples of past problems and their solutions. When they see a new situation, they quickly flip through their cheat sheet to find a similar past example and copy the solution.

How it works in the paper:

The Notebook: The AI reads thousands of past ship journeys. It writes down "past" (where the ship was) and "future" (where it went next) into a digital notebook.
The Search: When a new ship appears, the AI looks at where the ship is now and searches its notebook for past ships that were in a similar spot.
The Prediction: It pulls the "future" parts of those similar past trips and uses them to guess where the new ship will go.

A Key Twist: The researchers realized that for ships, knowing where they are isn't enough. You also need to know how fast they are going (Speed Over Ground) and which way they are pointing (Course Over Ground). So, they added these details to the "cheat sheet" entries, making the memory much more accurate.

The Experiment: The Ocean Test

The team tested this system using real data from two busy areas:

The Gulf of Mexico (a huge, open area).
The New York Bight (a busy area near the coast).

They compared their "Memory AI" against six other top computer models. These other models were like the "standard students" who didn't have a cheat sheet. Some of them were very complex, trying to figure out how ships interact with each other, but they still lacked the specific memory of past individual trips.

The Results: A Clear Victory

The results were impressive. The Memory AI (called MANTRA) beat every other model by a huge margin.

The Score: In the Gulf of Mexico, the new model reduced prediction errors by up to 46% for the average path and 55% for the final destination compared to the best competitor.
The Consistency: It didn't just win once; it won every time, even when the prediction window was longer (predicting 30 minutes into the future).
Why it won: The paper suggests that because ships often repeat the same routes (cruising) but occasionally do rare maneuvers, having a notebook that stores exact past examples helps the AI recognize both the boring, repetitive trips and the rare, tricky ones.

What They Didn't Do (Important Limits)

The paper is very specific about what they did and didn't do:

They did not try to predict how ships talk to each other or avoid collisions directly. They focused purely on predicting the path based on the ship's own history.
They did not test this on real-time ship navigation systems yet. This was a computer simulation using historical data.
They did not include weather or ship size in this specific test (though they mention they might do that in the future).

The Bottom Line

This paper proves that giving an AI a "memory bank" of past ship journeys is a powerful way to predict where ships will go. It's like upgrading a GPS from just showing you the map to one that remembers exactly how you drove last time, making it much better at guessing your next move.

Technical Summary: AIS-Based Vessel Trajectory Prediction Using Memory-Augmented Neural Networks

Problem Statement
Accurate vessel trajectory prediction is critical for maritime situational awareness, collision avoidance, and route optimization. While deep learning approaches have been widely applied to this task using Automatic Identification System (AIS) data, existing methods often struggle to capture the full spectrum of vessel behaviors. Vessel motion is characterized by frequent, repetitive cruising patterns interspersed with rare but diverse maneuvering behaviors. Standard sequence models may fail to explicitly retain specific instance-level trajectory patterns, particularly when maneuvering events are infrequent. This paper investigates whether memory-augmented neural networks (MANNs), which have shown success in pedestrian and road-vehicle prediction, can effectively address these challenges in the maritime domain.

Methodology
The authors propose an empirical investigation using MANTRA (Memory-Augmented Neural Network for Trajectory Prediction), a seminal MANN architecture, adapted for AIS data. The core methodology involves an external memory mechanism that stores and retrieves past-future encoding pairs.

Data Representation: The model utilizes spatial positions (latitude, longitude) alongside dynamic navigational attributes: Speed Over Ground (SOG) and Course Over Ground (COG). SOG and COG are incorporated as input covariates ( $x_t$ ) alongside position ( $y_t$ ), acknowledging their importance in vessel trajectory prediction. Missing AIS data is handled via interpolation, and a binary indicator is used to flag originally unobserved positions.
Memory Construction: An autoencoder framework is employed consisting of a past encoder ( $g$ $g$ ), a future encoder ( $h$ $h$ ), and a decoder ( $d$ $d$ ).
- The past encoder processes observed history $[y_P, x_P]$ to generate a past encoding ( $\pi$ ).
- The future encoder processes the ground-truth future $[y_F, x_F]$ to generate a future encoding ( $\delta$ ).
- These pairs are used to train the system to reconstruct future trajectories.
Memory Controller: A controller ( $c$ ) manages the external memory $\Omega$ . It decides whether to write a new trajectory segment $(\pi, \delta)$ into memory based on a "novelty" score. This score is derived from the prediction error ( $e_i$ ) of the current segment; if the model fails to predict a segment accurately using existing memory, the controller is trained to write that segment into memory to capture the missing pattern.
Prediction Phase: During inference, the model computes the past encoding for a new query trajectory. It retrieves the top- $K$ most similar memory items from $\Omega$ using cosine similarity. These retrieved future encodings are decoded to generate $K$ multi-modal trajectory predictions, accounting for the inherent uncertainty in vessel movements.

Key Contributions

First Empirical Investigation: To the authors' knowledge, this is the first study to apply memory-based trajectory prediction specifically to AIS-based vessel trajectory data.
Performance Gains: The study demonstrates that incorporating an external memory yields substantial performance improvements over deep learning baselines that lack memory mechanisms.
Handling Multi-Modality: The approach effectively captures both common cruising patterns and rare maneuvering behaviors by explicitly storing instance-level encodings, outperforming models that rely solely on temporal dependencies or inter-agent interactions.

Experimental Results
Experiments were conducted on AIS data from two regions: the Gulf of Mexico and the New York Bight. The proposed MANTRA model was compared against six baselines, including STGAT, Social-STGCNN, TransformerTF, AgentFormer, Social-Implicit, and TUTR.

Metrics: Performance was evaluated using Average Displacement Error (ADE) and Final Displacement Error (FDE) across prediction horizons of 10, 20, and 30 minutes.
Quantitative Outcomes: MANTRA consistently outperformed all baselines.
- Gulf of Mexico: Compared to the best-performing baseline, MANTRA reduced ADE by up to 46.4% and FDE by up to 54.7% (at 10-minute, 20-minute, and 30-minute horizons).
- New York Bight: Improvements reached up to 33.3% in ADE and 27.7% in FDE.
Stability: MANTRA exhibited the smallest standard errors across all settings, indicating more stable performance across random seeds compared to the baselines.
Qualitative Analysis: Visualizations confirmed that MANTRA closely matched ground-truth trajectories for both cruising and maneuvering scenarios.

Significance and Claims
The paper claims that memory-based approaches represent a powerful paradigm for vessel trajectory prediction. By explicitly retaining instance-level trajectory representations, the model can learn from specific past instances rather than relying solely on generalized statistical patterns. This capability is particularly valuable for maritime operations where rare maneuvering events are critical for safety but underrepresented in standard training distributions.

The authors conclude that while the current study focuses on spatial and dynamic navigational data, future work could integrate high-dimensional contextual variables (e.g., vessel type, weather, waterway geometry) and methods for quantifying prediction uncertainty. The study establishes a strong empirical foundation for the use of external memory in maritime AI applications.