Guardian Sentry · Spill drift modelling

Where would the oil go?

When Sentry records a release observation on a wreck, the next question is operational: if oil is in the water, where does it drift, what does it reach, and how long is there to respond? Guardian answers it with a decoupled drift-modelling engine: an independent service Guardian dispatches an incident to and consumes the assessment back from, with no shared datastore in either direction. When a detection escalates to survey, the attribution run dispatches automatically; the live Solomon coupling below shows the detection-to-attribution pipeline already running.

01 · Observe

Detect

Sentinel-1 SAR via SkyTruth Cerulean surfaces a slick and correlates it to candidate source wrecks by proximity, fuel and depth.

02 · Attribute · live

Rank the source wrecks

Each observation refines which wreck is the likely source — by proximity, fuel, depth and whether the wreck's modelled drift actually reaches the slick. This is attribution, not scoring: it does not move any wreck's WERP score. The one exception is a recurring slick with a single candidate — a wreck-specific release signal. See it live →

03 · Model · auto on escalation

Project the drift

The incident is dispatched to the drift engine, which returns a fate budget, a modelled footprint and the environmental exposure. Dispatch fires automatically when a slick goes under survey; other detections are dispatched by the analyst.

What the engine returns

An assessment bundle per incident

For any dispatched incident — a slick-to-source correlation or an on-demand wreck query — the decoupled engine returns a structured decision-support bundle. Forward drift from the source, a mass-balance fate budget, the surface footprint and envelope, and the environmental exposure along the drift path.

Forward drift

Source point, centroid track and modelled envelope over the response window.

Fate budget

Mass partitioned across surface, evaporation, shoreline and water-column.

Surface footprint

Swept area and net displacement, with multi-model spread as the uncertainty band.

Environmental exposure

Sensitivity of the receptors the drift actually reaches, under the Maximum Hazard Principle.

Metocean models

CMEMS global analysis-forecast (primary) and HYCOM ESPC-D-V02 as the second ocean model; Open-Meteo GFS 10 m wind. Waves and Stokes drift not yet coupled.

What the model is honest about

Release unconfirmed

The assessment models a hypothetical release; no release is asserted and the cargo grade is assumed, not confirmed.

Environmental layers are placeholders

The sensitivity overlay uses approximate layers pending authoritative boundaries (WDPA, IUCN, CBD). Structured, not authoritative.

Physics still being coupled

Waves and Stokes drift are not yet in the model (WAVERYS staging pending), so drift is currents plus wind only.

Independently benchmarked

Against an established reference (Marcast / SpillConsult) on identical inputs the engine currently under-drifts and under-entrains. These are calibration signals, not errors. Comparison (PDF) →

In the Solomon case

For the Iron Bottom Sound trio (Furutaka, Fubuki, Kirishima), Cerulean has recorded a recurring release-observation signal and Sentry's attribution is live. Because the slick has three plausible source wrecks, it stays attribution-only: the recurring signal sharpens which of the three is the likely source — by proximity, fuel, depth and drift-funnel reachability — without moving any wreck's WERP score. A recurring signal only raises release probability where a single wreck is the sole candidate.

The automated step — dispatching a backward-attribution run to the drift engine so it scores the candidate sources independently, alongside the correlation's own ranking — is live for slicks that escalate to survey; the two rankings are stored side by side, never blended. Detections below that threshold are dispatched manually by the analyst.

Open the regional monitoring picture →

The drift engine is consumed as an independent decoupled service (incident specification in, assessment bundle out; no cross-imports, no shared datastore). Its assessments are structured decision-support, not a definitive forecast or attribution: primary responders are sovereign. Engine outputs generated using E.U. Copernicus Marine Service Information and the GEBCO Compilation Group grid; slick detection from SkyTruth Cerulean under CC-BY-4.0.