Context-Enriched Natural Language Descriptions of Vessel Trajectories

This paper proposes a context-aware framework that transforms raw AIS vessel trajectory data into structured, semantically enriched representations annotated with multi-source contextual information, thereby enabling the generation of controlled natural language descriptions by LLMs to facilitate higher-level maritime reasoning and analytics.

The Gist

Imagine you have a massive, chaotic diary written by a ship. Every few seconds, it scribbles down a single coordinate: "I am here. Now I am here. Now I am here." This is what AIS data (Automatic Identification System) looks like. It's accurate, but it's boring, noisy, and completely lacks context. It's like trying to understand a movie by reading a list of every single frame number without knowing who the characters are or what the plot is.

This paper proposes a clever way to turn that boring list of coordinates into a rich, readable story that both humans and computers can understand. Here is how they do it, broken down into simple steps:

1. The "Smart Editor" (Cleaning and Segmenting)

First, the system acts like a smart film editor. Raw ship data is messy; sometimes the signal drops, sometimes the ship drifts, and sometimes it just sits still.

• The Analogy: Imagine watching a shaky home video of a road trip. The "Smart Editor" cuts out the blurry parts, the static, and the moments where the camera was just sitting on the dashboard.
• What it does: It groups the raw points into logical chunks called Episodes. Instead of "Point A, Point B, Point C," it says: "The ship stopped at the port," "The ship turned sharply," or "The ship sailed smoothly for an hour."

2. The "Travel Agent" (Adding Context)

Next, the system acts like a knowledgeable travel agent who knows the map better than anyone. It takes those "Episodes" and adds context.

• The Analogy: If you tell a friend, "I drove 50 miles," they might ask, "Where? Was it raining? Did you pass a famous mountain?"
• What it does: The system looks up the map and the weather. It adds details like: "The ship was sailing near a dangerous reef," "It was passing through a narrow strait," or "It was battling 40mph winds." It turns a dry math problem into a dramatic maritime adventure.

3. The "Storyteller" (Using AI to Write the Narrative)

This is the magic part. The team feeds this structured, context-rich data into a Large Language Model (LLM)—the same kind of AI that writes essays or chat messages.

• The Analogy: You hand the AI a stack of bullet points and a map, and say, "Write a travel blog post about this trip."
• What it does: The AI doesn't just list numbers; it writes a natural language story.
  • Raw Data: Lat: 55.1, Lon: 12.3, Speed: 14, Wind: 4B
  • AI Story: "The vessel sailed southwest off the coast of Lolland, cruising at 14 knots under a moderate westerly breeze, before making a sharp turn..."

Why Does This Matter?

The paper tested this with real ships and found some interesting things:

• Bigger is Better (for math): The biggest, most powerful AI models were the best at writing the story and getting the math (distance, speed, time) right. Smaller models often wrote great stories but got the numbers wrong (like saying a 2-hour trip took 20 hours).
• Safety First: By turning raw data into stories, port authorities and coast guards can quickly understand what a ship is doing without needing to be data scientists. If a ship takes a weird detour, the AI story can highlight it immediately: "The ship deviated from its usual route to avoid a storm."

The Big Picture

Think of this research as building a translator between the cold, hard language of machines (coordinates and timestamps) and the warm, rich language of humans (stories and context).

By doing this, they aren't just making pretty text; they are creating a foundation for smarter maritime safety. It allows computers to "read" the ocean's history, predict where ships might go next, and explain why they are doing it, all in plain English.

Technical Summary

1. Problem Statement

The paper addresses the challenge of transforming raw Automatic Identification System (AIS) data into structured, semantically enriched representations that are interpretable by both humans and machine reasoning systems (specifically Large Language Models or LLMs).

• Limitations of Raw AIS: While AIS provides precise spatiotemporal data (position, course, speed), the raw data is inherently noisy, occasionally sparse, and lacks explicit semantic annotation. It consists of unstructured tuples devoid of context (e.g., geographic features, weather, regulatory zones).
• LLM Integration Barrier: Direct interaction between raw AIS sequences and modern LLMs is hindered because LLMs struggle to infer navigational intent, operational patterns, or regulatory compliance from unprocessed coordinate streams without domain-specific grounding.

2. Methodology

The authors propose a context-aware trajectory abstraction framework that processes raw AIS data into "semantic trajectories" before feeding them to LLMs. The workflow consists of three main stages:

A. Trajectory Annotation & Segmentation

1. Mobility Event Detection: The framework uses a trajectory compression algorithm to annotate raw AIS points with specific mobility events:
   • Stops: Vessel stationary ($v < 0.5$ knots).
   • Turns: Significant heading deviation ($>5^\circ$).
   • Slow Motion: Consistent low speed ($<1$ knot) while moving.
   • Speed Changes: Acceleration/deceleration exceeding a threshold.
   • Communication Gaps: Missing AIS signals (distinguished between short gaps within a trip and long gaps splitting trips).
2. Segmentation:
   • Trips: Trajectories are split into distinct "trips" based on consecutive stops or long communication gaps ($\delta > 3$ hours).
   • Episodes: Each trip is further segmented into "episodes" representing specific motion patterns (e.g., Sailing, Turning, Maneuvering, Stopped).

B. Context Enrichment

Each episode is augmented with multi-source contextual data to provide semantic grounding:

• Geospatial: Proximity to ports, capes, straits, and islands (from OpenStreetMap, OpenSeaMap).
• Maritime: Traffic separation schemes, shipping lanes, and anchorage zones.
• Meteorological: Wind force (Beaufort scale) and direction (from Copernicus/NOAA).
• Bathymetry: Seabed depth (from GEBCO) to explain course deviations due to shallow waters.

C. LLM-Based Narrative Generation

The enriched semantic trajectory (structured as JSON) is fed into an LLM via a specialized prompt.

• Input: A JSON array containing episode statistics (duration, distance, speed, heading) and contextual tags (placemarks, weather).
• Output:
  1. Plain Text Narrative: A factual, step-by-step description of the voyage (e.g., "The vessel sailed south-west off Lolland...").
  2. Structured Statistics: A JSON summary containing total distance, duration, origin/destination ports, and adverse weather conditions.

3. Key Contributions

1. Semantic Trajectory Framework: Development of an open-source framework that converts noisy AIS data into structured "trips" and "episodes" annotated with mobility events.
2. Multi-Source Context Integration: Systematic enrichment of trajectory segments with heterogeneous data (geographic, maritime, meteorological, and bathymetric) to explain vessel behavior.
3. LLM-Driven Narrative Synthesis: A novel pipeline using LLMs to translate structured semantic trajectories into human-readable, factual natural language descriptions, bridging the gap between raw data and high-level reasoning.
4. Empirical Evaluation: A comprehensive study using real-world AIS data (Danish passenger ships) to evaluate the quality, accuracy, and robustness of descriptions generated by various LLMs (ranging from 8B to 120B parameters).

4. Experimental Results

The study utilized 460 trips (212,203 raw AIS locations) and tested five LLMs (Llama-3.1/3.3, Qwen, GPT-OSS).

• Context Enrichment:
  • Over 21% of episodes involved turns or maneuvers.
  • 100% of "Stopped" episodes were successfully linked to port names.
  • Bathymetry data was available for the vast majority of sailing and turning legs.
• Performance (Response Time):
  • Smaller models (8B) generated responses in ~1 second; larger models (120B) took ~2.4 seconds. All are suitable for near-real-time applications.
• Quantitative Accuracy (Statistics):
  • Large models (e.g., GPT-OSS-120B) significantly outperformed smaller ones in estimating total trip duration and distance, with mean deviations of ~1.5% and ~0.9% respectively.
  • Small models (e.g., Llama-3.1-8B) showed poor calculation capabilities, with mean duration deviations exceeding 113% and maximum deviations reaching nearly 19,000%.
• Qualitative Assessment (LLM-as-a-Judge):
  • Evaluated on Relevance, Faithfulness (grounded in facts), and Correctness.
  • GPT-OSS-120B achieved the highest scores (4.91/5.0 Relevance, 4.96/5.0 Faithfulness, 4.69/5.0 Correctness).
  • Smaller models struggled with Correctness, often hallucinating speed or duration values due to limited arithmetic reasoning capabilities.

5. Significance and Applications

• Explainable AI (XAI): The generated narratives provide transparent reasoning for vessel behavior (e.g., explaining a detour as a response to a storm or shallow water), crucial for trust in maritime AI.
• Downstream Analytics: The structured representations enable better trajectory prediction, anomaly detection (e.g., identifying unauthorized detours), and data imputation for missing AIS segments.
• Operational Utility:
  • Safety: Assists in contextualizing accidents or near-misses.
  • Compliance: Automatically generates voyage reports to verify adherence to charter agreements and regulatory zones.
  • Navigation: Provides human-readable route summaries similar to official sailing directions.

Conclusion

The paper demonstrates that transforming raw AIS data into context-enriched semantic trajectories is a critical preprocessing step for effectively leveraging LLMs in the maritime domain. While LLMs show great promise in synthesizing these narratives, the study highlights that larger models are essential for maintaining factual accuracy in quantitative metrics, whereas smaller models may require human oversight for calculation-heavy tasks. This framework lays the groundwork for more intelligent, explainable, and safety-oriented maritime monitoring systems.