Finding the ships that go dark.
When a vessel switches off its AIS transponder, it disappears from the usual picture. Planetar learns to recognise that same vessel in satellite radar, harbour cameras, underwater microphones and other sensors, so it can be followed after the beacon goes silent.
How it learns
-
Watch while the beacon is on
While a ship broadcasts AIS, its name and position tell us which radar returns, camera frames and sounds belong to it. Every hour of normal traffic gives the model free training examples, with nobody labelling data by hand.
-
Learn one shared picture
Each sensor gets its own encoder, trained so that observations of the same vessel land close together whichever sensor made them.
-
Re-identify when it goes dark
Once the beacon goes silent, new sightings are matched back to the vessel. Each match has a calibrated confidence and shows the evidence it rests on, and an analyst confirms or rejects it.
Sensors
The ingest services in bold are built and running on the shared message bus. The learned model that fuses them is the current research.
- AIS
- Live ship positions from coastal receivers
- Satellite radar
- Sentinel-1 radar scenes, with ship detection
- Cameras
- Public harbour webcams, with vessel detection
- Hydrophones
- Underwater audio, classified for vessel sound
- Radio emissions
- Planned
- Text reports
- Planned: notices to mariners and open reporting
Live sites
The analyst workspace on live AIS traffic around Victoria, BC: a channel for each vessel, a map, the entity graph, and a trace of every message from sensor to screen.
Where are the whales around Victoria right now, and why do we think so? It estimates orca positions from how whale-watching boats move in live AIS, and every estimate links to the data behind it.
Open source
The Planetar services are public under the GNU GPL v3. The wire format, zmesg, is Apache-2.0 so anyone can speak it.
- planetar-broker the message bus, in C
- planetar-ontology the vessel and entity graph
- planetar-ui the analyst workspace
- planetar-ais AIS ingest
- planetar-sat satellite radar
- planetar-eo cameras
- planetar-acoustic hydrophones
About
Planetar is a project of Zax Analytics, an independent Canadian company in Victoria, British Columbia. Steven Ness, the company's founder and CEO, builds it. He brings years of machine-learning work on large unlabelled audio archives, including killer-whale calls from Ocean Networks Canada hydrophones.