Beyond geofencing: Behavior detection using AIS

Raphael Sturgis, Valentin Emiya, Basile Couëtoux, Pierre Garreau

TL;DR

This paper tackles the challenge of accurately detecting vessel behaviors (like underway, moored, or drifting) from Automatic Identification System (AIS) data, which is often plagued by geographical biases. They propose two novel methods – data augmentation and unbiased feature engineering – to create models that generalize well across different regions, outperforming traditional geofencing approaches.

What Problem Does This Paper Address?

Geofencing, a common method for detecting vessel behavior, relies on predefined geographical areas (polygons) to infer behavior. While effective in some cases, it has significant drawbacks: it's costly and time-consuming to create and maintain these polygons, and it struggles with dynamic situations like port congestion or behaviors that aren't tied to specific locations (e.g., drifting). Furthermore, geofencing is limited in its ability to capture nuanced behaviors. AIS data, while rich, contains geographical biases that hinder the generalization of machine learning models trained in one region to another. This paper aims to overcome these limitations by developing methods to detect vessel behavior directly from AIS trajectories, ensuring models are robust and perform well globally, even in areas with limited or no labeled data.

Methods

The authors propose two main strategies to combat geographical biases in AIS data for behavior detection:

  • Data Augmentation: They introduce a novel technique to augment AIS trajectories by applying rotations and translations on the surface of a sphere. This process aims to create new, plausible trajectory variations that help models learn to generalize better by exposing them to a wider range of scenarios without altering the underlying behavior. This is achieved using geometrical tools like quaternions for accurate rotations.
  • Unbiased Feature Engineering: This approach focuses on creating new features from the raw AIS data that are less susceptible to geographical influences. They achieve this by using operations like local derivatives and standard deviations on features like speed, course, and heading. The goal is to extract information about the change and spread of these features, which are less dependent on specific geographical contexts compared to raw values.

They then evaluate these strategies using various machine learning models, including Decision Trees (DT), Random Forests (RF), Hidden Markov Models (HMM), Long Short-Term Memory networks (LSTM), and Transformer Networks (TN).

Datasets

The study utilizes two datasets of AIS data for container vessels:

  • North-East (NE) Dataset: Collected from the north-eastern United States territorial waters, covering January 1 to March 31, 2021. This dataset was used for both training and testing.
  • Los Angeles & Long Beach (LAX) Dataset: Collected from the ports of Los Angeles and Long Beach, California, covering the same period. This dataset was used exclusively for testing to evaluate inter-regional generalization.

Both datasets were manually labeled into four classes: Underway, Moored, At Anchor, and Drifting. The paper notes that these datasets are unbalanced.

Key Insights and Ideas

  • SOG is King (for generalization): Speed Over Ground (SOG) is identified as the most geographically unbiased feature. Models relying solely on SOG showed significantly better generalization across different geographical areas compared to models using other features like course, heading, or location.
  • Sequential Models are More Resilient: Models designed for sequential data (LSTM and TN) demonstrated better resilience to geographical biases compared to single-point classifiers like DT and RF when using base features.
  • Unbiased Features are Crucial for Inter-Area Performance: The engineered unbiased features significantly improved cross-area performance, particularly for F1 scores, indicating a more robust understanding of vessel behavior regardless of location.
  • Data Augmentation Offers Modest Gains: While data augmentation showed some improvement, especially for Transformer Networks, its impact was less pronounced than unbiased feature engineering, particularly when testing on data from a different region.
  • Navigation Status is Unreliable: The AIS-provided "Navigation Status" field, often used as a baseline, proved to be a poor indicator of actual vessel behavior, especially for moored and drifting states, highlighting the need for more sophisticated detection methods.

Key Results

  • Unbiased Features Outperform: Models trained with unbiased features consistently achieved higher accuracy and Macro F1 scores across both the NE and LAX datasets compared to models trained on base features or with data augmentation.
  • Random Forest with Unbiased Features is Top Performer: The Random Forest model, when trained with unbiased features, achieved the highest accuracy and Macro F1 scores on both datasets, demonstrating very stable performance.
  • Transformer Networks Show Promise: Transformer Networks also showed strong performance, especially when combined with unbiased features, and exhibited less variance than LSTMs when using base features.
  • Significant Accuracy Drop Without Bias Mitigation: Configurations using features like cog, heading, and latitude/longitude without bias mitigation showed a substantial drop in accuracy when tested on a different geographical area compared to the training area.
  • Event Detection Challenges: Detecting "At Anchor" events remained a challenge for most models, with performance often not exceeding the unreliable Navigation Status baseline.

Limitations and Future Work

  • Data Augmentation Limits: The study found that data augmentation provided no or minor improvements on the same-area dataset (NE) and more significant gains for TNs on the cross-area dataset (LAX), but the augmentation factor didn't show a clear optimal value beyond a certain point.
  • Event Detection for Rare Behaviors: The paper acknowledges that detecting rare behaviors like piracy or accidents remains a challenge and suggests unsupervised learning approaches for such scenarios.
  • Future Work: The authors suggest exploring the dynamical creation of semantic maps for automatic geofencing, generating polygons in unlabeled areas, and studying port congestion. They also propose integrating additional information like meteorological data or focusing on other fishing behaviors. The potential for using unsupervised learning on large amounts of unlabeled AIS data for fine-tuning is also highlighted.

Summary updated

BibTeX
@inproceedings{sturgis2024beyond,
  author = {Raphael Sturgis and Valentin Emiya and Basile Couëtoux and Pierre Garreau},
  title = {{Beyond geofencing: Behavior detection using AIS}},
  booktitle = {Ocean},
  year = {2024},
  doi = {10.1016/j.oceaneng.2023.116630},
}
Figure