SeaVigil

Methodology and data sources

Back to the map

SeaVigil is the explanation-and-evidence layer for illegal-fishing behaviors in protected and national waters. It does not try to re-detect fishing at global scale (that is Global Fishing Watch's industrial-scale job). It sits on top of vessel-monitoring signals and turns each flag into a per-incident, explainable, auditable dossier that an enforcement officer can read, defend, and act on, running on a single laptop with no cloud account.

Illegal, unreported and unregulated (IUU) fishing is estimated at 11 to 26 million tonnes of catch per year, worth roughly USD 10 to 23.5 billion (Agnew et al., 2009), and 72 to 76 percent of the world's industrial fishing vessels are not publicly tracked (Paolo et al., 2024). SeaVigil targets the last mile: explaining and evidencing the suspicious behaviors that monitoring surfaces.
  1. 1Inputs

    Satellites (radar and optical) and ships' own AIS radio signals

  2. 2Detect

    SeaVigil's models find vessels and judge whether the behaviour is suspect

  3. 3Flag

    Each suspicious vessel becomes a flagged incident in its sea zone

  4. 4Evidence

    A dossier: the reason, the place, and a satellite image of the boat

From satellites and ships' signals to an evidence-backed alert: the same path behind every flag on the map.

The terms, in plain words

Dark vessel. A boat that has switched off its AIS transponder. Radar still catches it as a blip, but it gives no name or position. Most of the world's industrial fishing vessels go untracked this way.

Going dark. A vessel that was broadcasting its position and then suddenly stops, often to hide where it is heading or what it is doing.

Transshipment. Two vessels meeting at sea so a fishing boat can hand its catch to a carrier, a common way to move illegal fish without ever returning to port.

EEZ (the dashed border). A country's Exclusive Economic Zone, its waters out to 200 nautical miles. A foreign boat fishing here without a licence is the classic illegal-fishing lead.

MPA (the solid border). A Marine Protected Area, where fishing is limited or banned outright. Any apparent fishing inside one is a red flag.

SAR and optical. The two satellite eyes. SAR is radar: it sees vessels through cloud and at night. Optical is a true-colour photo: sharper, but blocked by cloud.

What SeaVigil detects, and how

AIS apparent fishing inside a protected area

A RandomForest classifier scores each AIS position as fishing or not from vessel kinematics (speed, turning, distance from shore and port, time of day, rolling statistics), trained on Global Fishing Watch's openly published, hand-labelled AIS data (Kroodsma et al., 2018). Positions are joined by point-in-polygon to Marine Protected Area boundaries; runs of in-MPA fishing positions are segmented into incidents. Each incident's reason is the SHAP attribution (Lundberg and Lee, 2017) averaged over its positions, shown as a per-feature contribution chart.

Dark vessel (SAR) inside an EEZ or reserve

Vessels that broadcast no AIS ("dark") are the majority of industrial fishing activity. SeaVigil runs its own Sentinel-1 synthetic-aperture-radar detection (the open Allen Institute model, on the freshest Copernicus scene within hours of each satellite pass) and its own Sentinel-2 optical detection, and saves a true-colour satellite image chip of each detection (dark-fleet scale: Paolo et al., 2024). A detection not matched to any AIS broadcast, inside a national EEZ, is flagged dark; severity escalates inside a Marine Protected Area. The rationale is attribute-based (unmatched, inside-zone, length, recurrence), not SHAP, because a radar blip carries no movement track.

EEZ incursion, flag, and authorization

Every incident is tagged with the Exclusive Economic Zone it sits in (point-in-polygon against the global Marine Regions EEZ set) and the coastal state. When a vessel's flag differs from the EEZ sovereign it is marked foreign. A foreign vessel inside another state's EEZ is the canonical IUU lead, but on its own it is not proof: the vessel may be licensed.

So for every incident that carries a vessel identity (MMSI), SeaVigil looks the vessel up in the Global Fishing Watch vessel-identity registry to get its authoritative flag and the authorizations on record (RFMO and regional bodies such as FFA, WCPFC, IOTC, ICCAT, IATTC, CCSBT, CCAMLR), then grades the incident against the incident date: authorized (an authorization covering the date is on record), authorization lapsed (one existed but expired before the date), no authorization on record (foreign, nothing found), domestic flag, or no vessel identity (anonymized AIS labels and dark SAR cannot be checked). This turns a bare "foreign" flag into a graded, sourced status. The honest limit: national EEZ fishing licences are not public, so an empty authorization record means "no public record", not proof of illegality; the cleanest signal here is a lapsed or absent RFMO authorization for a foreign vessel.

Going dark and at-sea transshipment, worldwide

Going dark (a vessel that was broadcasting and then stops) and at-sea transshipment (two vessels meeting to move a catch) are detected live from the AIS stream, and are also drawn from Global Fishing Watch's published global datasets for the open ocean that terrestrial AIS cannot reach: intentional AIS disabling (Welch et al., 2022; a multi-hour gap well offshore) and transshipment (Miller et al., 2018; a fishing vessel and a carrier together for hours, below a few knots, away from port).

Model and confidence

The fishing classifier is evaluated on held-out vessels (grouped split, so no vessel appears in both training and test). On about 408,000 held-out positions it is well calibrated: a Brier score of 0.092, and isotonic recalibration only moves it to 0.088. In plain terms, a dossier score near 0.9 can be read honestly as "about 90 percent of positions scored this high really are fishing." Rule-based flags (dark SAR, EEZ incursion) carry no probability; their confidence is the explicit criteria they meet. The full reliability table is in results/calibration.json and the calibration statement appears in every AIS dossier.

Evidence and integrity

Each incident's downloadable dossier is a structured evidence package: the detection and its reason, the exact UTC times and coordinates, the vessel identity where known, the full data provenance below, and a SHA-256 integrity hash over the incident's canonical facts so a reviewer can recompute it and confirm the record was not altered. It carries the standing disclaimer that AIS and SAR evidence have coverage gaps and spoofing risks and that an alert is an inspection lead, not proof of illegality.

Data sources

DatasetReference / DOILicense
GFW labelled AIS training data Kroodsma et al., Science 2018. 10.1126/science.aao5646 CC BY 4.0
GFW Sentinel-1 SAR vessel detections Paolo et al., Nature 2024. 10.1038/s41586-023-06825-8 CC BY-NC 4.0
GFW vessel identity and authorizations Global Fishing Watch vessel-identity dataset (registry flag + RFMO/regional authorizations: FFA, WCPFC, IOTC, ICCAT, IATTC, CCSBT, CCAMLR). globalfishingwatch.org CC BY-NC 4.0
Marine Protected Areas (WDPA / WD-OECM) UNEP-WCMC and IUCN (2026), Protected Planet, June 2026. protectedplanet.net Non-commercial, display-only
Exclusive Economic Zones (v12) Flanders Marine Institute (2024), Maritime Boundaries Geodatabase. 10.14284/632 CC BY 4.0
Live AIS stream aisstream.io (near-real-time, Great Barrier Reef) Free tier
Coastline and ports (distance features) Natural Earth Public domain
Basemap (land and borders) Natural Earth, built into a local vector tileset so the map runs fully offline (no CDN) Public domain

Methods reference: AIS disabling, Welch et al., Science Advances 2022, 10.1126/sciadv.abq2109; transshipment, Miller et al., Frontiers in Marine Science 2018, 10.3389/fmars.2018.00240; explanations, Lundberg and Lee, NeurIPS 2017, arXiv:1705.07874.

Honest limitations