PhD Defence Peder Heiselberg

Principal supervisor:

Professor Ole Baltazar Andersen, DTU Space

Co-supervisor:

Associated professor John P. Merryman Boncori, DTU Space

Examiners:

Associated professor Bjørn Sand Jensen, DTU Space
Professor Tim McCarthy, University of Maynooth, Ireland
Professor Peter Nielsen, University of Aalborg, Denmark

Chairperson at defence:

Professor Jonathan Frank Kirby, DTU Space

Summary:

Larger ships are required to carry an AIS transponder, intended for collision avoidance. AIS allows the maritime traffic to be monitored. However, the system can be tampered with or simply turned off. In remote areas such as the Arctic, lack of coverage causes the AIS message to be lost. Ships that do not continuously transmit AIS messages are known as dark ships. Dark ships can be engaged in Illegal Unreported Unregulated (IUU) fishing, smuggling, oil spills and other illicit activities. These ships can not be monitored with AIS alone. Additional data sources are required. Satellite images provide an effective way of monitoring the maritime domain, including dark ships. Automatically extracting accurate information from the images is difficult and a major hindrance. It may be why satellite images has not seen widespread adoption for maritime monitoring, despite, in some cases, being freely available. A complete approach to dark ship detection in satellite images is presented in this thesis. Localizing ships in the images is not sufficient. AIS must also be correlated to validate a ship as dark. In the Arctic, there are numerous icebergs, which create an abundance of false alarms. Due to the huge remote area, reducing the number of false alarms is extremely important for search and rescue operations. This work presents a methodology for efficiently discriminating ships from icebergs. Satellites can monitor large areas, often trading resolution for coverage. The detection of a dark ship can trigger tasking of a high resolution satellite to take a closer look. There is also a latency from the satellite recording an image to it being available on the ground. Tasking a new satellite or intercepting the dark ship thus not only requires its location, but also its velocity and course. An investigation is conducted into utilizing a satellite image processing artifact to accurately estimate a ship’s velocity and course. One of the most employed tools in modern crime shows is the fingerprint. Drawing from it as inspiration, this thesis presents the first study into identifying ships in satellite images. The presented work uses state-of-the-art facial recognition algorithms and exploits the ship’s AIS to create a fingerprint database of ships. Dark ships can then be cross referenced against the database. In the aerospace domain, dark aircraft, are also known as Unidentified Flying Objects (UFO). Instead of AIS, aircraft transmit ADS-B. The setting is thus similar to that of dark ships, and many of the algorithms can be repurposed. This thesis presents a full deep learning framework for simultaneous detection and state (velocity, height, heading) estimation for aircraft in satellite images.