Encoding Temporal and Spatial Vessel Context using Self-Supervised Learning Model
Introduction
Maritime surveillance faces a significant "big data" challenge. While the Automatic Identification System (AIS) provides a continuous stream of vessel data—including location, speed, and heading—this data is frequently noisy, irregular, and largely unlabeled. Traditional surveillance methods often analyze a single ship’s trajectory in isolation. However, critical illegal activities such as smuggling, unauthorized fishing, or illicit ship-to-ship transfers often involve complex interactions between multiple vessels or specific environmental contexts.
Furthermore, the scarcity of labeled examples for confirmed illegal acts makes standard supervised machine learning prone to overfitting. To address these challenges, this work proposes a self-supervised representation learning model designed to capture both a vessel's individual movement history and its spatial relationship to surrounding traffic.
Methodology
The core of the proposed approach is a Self-Supervised Learning (SSL) framework. Instead of training directly on rare illegal activities, the model is first pre-trained on "pretext tasks"—synthetic puzzles derived from raw AIS data that require the model to understand maritime dynamics.
The architecture consists of three primary components:
- Temporal Encoding: A sequence-to-sequence architecture (such as an RNN or Transformer) processes a ship's message history into a fixed-width state vector.
- Spatial Context: To capture environmental interactions, the model aggregates the state vectors of the $N$ closest vessels. This allows the system to "see" the surrounding traffic density and behavior.
- Pretext Training: The model is pre-trained on six specific tasks, including predicting current heading, next position, distance traveled, and determining if a message is the logical successor in a sequence.
Once pre-trained, the encoder is "frozen," and a small output layer is added for fine-tuning on specific downstream "Main Tasks," such as AIS shutdown detection or identifying illegal rendezvous.
Experimental Results
The model was evaluated using AIS data streams, focusing on two critical downstream tasks: intentional AIS shutdown detection and illegal ship-to-ship rendezvous.
- AIS Shutdown Detection: The model achieved 92% accuracy. An ablation study revealed that removing the spatial context (ignoring neighboring vessels) dropped accuracy to 79.4%, confirming that surrounding traffic is a vital indicator for distinguishing between technical signal loss and intentional evasion.
- Illegal Rendezvous Detection: The system achieved 89.7% accuracy, demonstrating robust performance even with limited labeled training samples.
Discussion and Conclusion
The primary insight of this research is that vessel behavior cannot be fully interpreted in a vacuum. By integrating spatial context through SSL, the model learns the "physics" and behavioral norms of maritime traffic. This foundational knowledge allows for more accurate anomaly detection with minimal labeled data.
Future work will focus on expanding the suite of pretext tasks to improve feature coverage and applying the encoder to broader tasks such as vessel type classification and general trajectory anomaly detection. Refining the selection of pretext tasks remains a key challenge in optimizing domain-specific maritime representations.
Summary updated