Vessels anomaly detection with AIS data

Introduction

We takes inspiration from TREAD* (Traffic Route Extraction and Anomaly Detection)[1], that learns a statistical model for maritime traffic from AIS data in an unsupervised way

Dataset

National Oceanic and Atmospheric Administration: https://marinecadastre.gov/accessais

The information contained in the AIS data (generally) are of dynamic and static type

Area of interest: Hawaii

A bounding box is selected and it corresponds to the (chosen) specific area under surveillance

Waypoints identification

Waypoints (WPs): identify either stationary points (ports, offshore platforms, etc.), entry points and exit points.

WPs identification is based on DBSCAN (Density-Based Spatial Clustering of Applications with Noise) method

Note: The hyperparameters of DBSCAN are tuned manually.

Waypoints identification

Once the waypoints are learned, a route can be built by clustering the extracted vessel flows, which connect:

• Two ports
• Entry point and port
• Port and Exit point
• Entry point and Exit point (i.e., transit routes)

PoL learning (work in progress)

It is necessary to remove outliers from the routes to identify normal traffic. In TREAD, it is used the stochastic method Kernel Density Estimation (KDE).

References

[1] *Pallotta, Giuliana, Vespe, Michele and Bryan, Karna. "Vessel pattern knowledge discovery from AIS data: A framework for anomaly detection and route prediction." Entropy 15.6 (2013): 2218-2245.