Models

Ghost Fleet adds satellite radar to the tracking data. This page says what each model does, what it was measured against, and where it falls short. No number here is proof of wrongdoing.

Loading the model report…

1. Radar detections from Global Fishing Watch

Global Fishing Watch detects vessels in Copernicus Sentinel-1 radar images and matches each detection to a ship broadcasting AIS where it can. We collect those detections for the corridors where our tankers spend time, and count the ones matched to each tanker's IMO. This is observed data from GFW's own models, not ours.

2. Our vessel detector

A small convolutional network (about 120,000 parameters) reads a 1.6 km radar chip and says whether it holds a vessel, and where. It is compared with a classical CFAR detector, the standard non-learning method.

Split by radar scene: no image in the test set was seen in training. The labels are GFW's detections, so these figures measure agreement with GFW's detector, not ground truth. GFW misses some small vessels, so a few "empty" chips may hold one.

3. Matching a tanker to a radar target

When a Sentinel-1 image was taken while AIS showed one of our tankers loitering, we run the detector on the surrounding 3.2 km and score each target by detector confidence and distance, including the chance that the ship was not seen. GFW's detections are not used to make the match; GFW's own AIS match is used to check it.

4. Possible side-by-side activity

Where a second target sits within 250 m of a matched tanker, the image goes to review. Anchorage queues, bunkering, tugs and rafted ships look the same from orbit, so a confirmed pair is "possible side-by-side activity", never a ship-to-ship transfer.

5. Loaded or empty?

Free data has no draught readings. We tested whether radar alone can tell likely-laden from likely-ballast tankers, using voyage context as weak labels (after a Russian export terminal: likely laden; heading to one: likely ballast), with every vessel's images kept out of its own training fold. The pass mark was fixed before the test.

6. Independent test: xView3-SAR

Sections 2 and 3 are measured against Global Fishing Watch, whose detections also trained the detector. xView3-SAR (DIU) is an independent reference: vessels labelled by analysts and AIS on Sentinel-1 scenes from 2020, in other seas. We ran the same detector, unchanged, on chips centred on xView3's high- and medium-confidence vessels and on open water at least 2 km from any xView3 detection.