Beyond AIS: how agents fuse satellite, weather and registry feeds

AIS gets the attention because it is the feed you can see on a map. But it is one feed, it is cooperative, and it is spoofable — and a 2024 Nature study of two petabytes of satellite imagery found that roughly 75% of industrial fishing vessels and over a quarter of transport and energy vessel activity are absent from public AIS. Any serious maritime question therefore lives in the space between feeds. The interesting problem was never AIS; it was fusion.

Fusion used to be a project. Now it is a call.

Historically, combining feeds meant a data-engineering programme: license SAR, license weather, license registries, build pipelines to normalise and join them, keep them fresh, and only then let an analyst ask a question. The fusion was baked in advance, which meant it was expensive, brittle, and limited to the questions someone anticipated.

A tool-using agent changes the economics. Instead of pre-joining everything, the agent calls each feed on demand, in the course of answering a specific question. The fusion happens at investigation time, driven by what the question needs — not at ingestion time, driven by what an engineer guessed. That makes multi-feed analysis something you do, not something you build.

What a fused investigation looks like

Take the canonical dark-vessel question — “what was this tanker doing during its AIS gap?” A single feed cannot answer it. An agent fusing feeds can:

1. AIS (cooperative): pull the vessel’s track and isolate the gap window, normalised against expected satellite coverage so a thin-coverage patch is not mistaken for evasion.
2. SAR (non-cooperative): request Copernicus Sentinel-1 imagery for that time and place — radar sees hulls in any weather, day or night — and run detection to find vessels present but silent.
3. Association: match radar detections to AIS returns; a detection with no matching transponder is a candidate dark ship. Optical imagery (Sentinel-2, or commercial ICEYE/Planet) confirms in clear conditions.
4. Metocean (environmental): add weather and current context — was the vessel loitering, or holding station in heavy seas? Context separates intent from circumstance.
5. Registry (reference): resolve identity across MMSI and IMO and pull ownership from Equasis, Clarksons and S&P, then check any co-located vessel — is there a shared manager, a known shell structure, a sanctions link?

The output is one explained conclusion with evidence from every layer, not five disconnected exports. This is the pipeline behind our maritime domain awareness and the sanctions work in detecting a vessel that has turned off its AIS.

Different feeds, different questions

Fusion is not only for dark vessels. The same on-demand orchestration serves the whole value chain:

• Emissions: fuse AIS activity with vessel fuel curves and calibrate against verified EU MRV data to produce MRV-grade CO₂ estimates — the basis of our decarbonization work.
• ETA and routing: fuse AIS voyage history with CMEMS currents, WaveWatch III waves and ECMWF wind, bounded by GEBCO bathymetry, for weather-aware arrival prediction.
• Flow intelligence: fuse port-call time series with UN Comtrade trade data for demand and floating-storage signals.

The feed list is long — the full catalogue is on our data & methods page — but the pattern is always the same: the agent reaches for whichever sources the question needs.

Retrieval over the unstructured, too

Not every feed is a clean API. A lot of decisive maritime context is text: sanctions designations, Port State Control records, notices to mariners, class-society bulletins, news. Agents handle this through retrieval — searching a corpus of registry and regulatory documents and grounding the answer in what they find, with citations. That is what lets an agent say “this ownership structure matches a designation published last week” instead of relying on a stale, pre-built table.

The interoperability that makes it work — and who it favours

On-demand fusion depends on the agent being able to reach the feeds. Standard tool interfaces — the Model Context Protocol and similar — let an agent call AIS, satellite, weather and registry sources directly, alongside a client’s own systems, without a bespoke integration for each. That has a strategic consequence: it favours vendor-neutral orchestration. An agent that can call any feed answers the question with the best available source, rather than only the data that happens to live inside one platform. It is the technical foundation of the independent analytics layer we argue for — and the reason fusion, not any single feed, is the durable advantage.

The honest caveats

Multi-feed fusion is powerful and imperfect. Satellite tasking and revisit limit when you can corroborate; imagery resolution limits what you can classify; registry data is incomplete and sometimes deliberately obscured; latency and coverage vary by feed. A good agent surfaces these limits rather than papering over them — “no SAR pass was available within the gap window” is a more useful answer than a false certainty. Fusion widens the picture; it does not make it complete, and saying so is part of doing it well.