From open source to operational: scaling a live OSINT maritime picture with proprietary data and fine-tuned models

Most maritime-intelligence vendors show you a polished dashboard and ask you to trust the data behind it. We did the opposite: we built the open-source version in public, on free data, and published exactly what it can and cannot see. live.marineaware.com is a live maritime domain awareness (MDA) dashboard that scores 12 global chokepoints — Hormuz, Bab-el-Mandeb, the Taiwan Strait, the Baltic cable zone and more — every day, fusing four open signal layers into one picture. This piece is about what that demonstrator proves, where open data hits its ceiling, and how the identical fusion engine scales with proprietary data and fine-tuned models.

What we built on open data alone

The live dashboard fuses, per chokepoint, entirely from free sources:

• Live AIS — vessel positions, names and speeds streamed from open AIS, plotted on a per-hotspot map with under-way, slow and stationary states.
• Dated OSINT reporting — recent, date-stamped news and open reporting per zone, deduplicated and placed on a 21-day timeline so a coverage spike is visible, not just asserted.
• Metocean — real wind, gusts, wave and swell height and visibility at each box, resolving to a transit read.
• Satellite — a daily open true-color snapshot (NASA GIBS) for context.

A deterministic model blends these into a 0–100 risk score, and the whole site is static, rebuilt daily. It is a genuine, working intelligence product — and it is honest about being the free tier. That honesty is the point: it makes the scale-up legible.

Where open data hits its ceiling

Three limits show up immediately, and they are the same three that separate a demo from an operational system.

AIS is rate-limited and cooperative. Free AIS throttles hard across many areas at once, so some boxes show a handful of vessels when hundreds are really there. And AIS is self-reported — it answers what cooperative vessels choose to broadcast. A 2024 Nature study of two petabytes of satellite imagery found roughly 75% of industrial fishing vessel activity, and over a quarter of transport and energy vessel activity, absent from public AIS. The interesting vessels are frequently the silent ones.

Open satellite is coarse and slow. Copernicus Sentinel-1 SAR and Sentinel-2 optical are extraordinary public goods, but revisit is measured in days and resolution in tens of metres. A daily true-color tile is context, not evidence — far too coarse to count ships or read a wake. Open imagery samples the ocean; it does not watch it.

Open reporting is attention, not ground truth. News volume tells you where the world is looking, which is useful, but it lags events and can spike on anniversaries. It corroborates; it cannot originate the finding.

None of this makes the open picture worthless — it makes it a baseline. Knowing precisely where the baseline ends is what lets you specify what proprietary data has to add.

Where proprietary data changes the picture

The scale-up does not change the method — fusion across feeds — it changes the grade of each feed.

• Licensed global AIS (Kpler, Spire and satellite constellations) replaces the throttled open stream with full-resolution, near-real-time coverage over open ocean, so a vessel count is a count, not a sample, and an AIS gap is a real gap rather than a coverage artefact.
• High-resolution commercial SAR (ICEYE, Umbra, Capella) replaces multi-day open revisit with persistent, all-weather tasking. A radar detection with no matching AIS return is the canonical dark-ship signal; at 16–25 cm and high revisit, you can not only find the silent vessel but classify and re-observe it.
• Commercial optical and hyperspectral adds daylight confirmation and material discrimination the free tier cannot reach.

The public dashboard already demonstrates the fusion pattern. Proprietary data is what makes each layer operational — and it is exactly the upgrade path we build for clients, from private areas of interest to alerting on licensed feeds.

Multispectral and hyperspectral: seeing what radar and RGB miss

SAR sees shape and presence in any weather; RGB optical shows a scene a human recognizes. Neither reads material. Multispectral and hyperspectral sensors do — they sample dozens to hundreds of narrow spectral bands, so the data carries chemistry, not just geometry.

For maritime work that unlocks signatures a hull outline never will: oil sheens and bilge discharge on the water, gas flaring intensity as a proxy for energy export, chlorophyll and turbidity that shape fishing behaviour, and spectral cues that help separate vessel types and cargo states. Planet’s Tanager hyperspectral satellites capture 426 bands across the 380–2500 nm range and already ship methane-plume products; Pixxel is standing up a commercial hyperspectral constellation. Fused with SAR detection and AIS identity, a multispectral pass turns “there is a vessel here” into “there is this kind of vessel, doing this, leaving that in the water.”

This is a data type the open tier essentially does not offer at useful cadence — and one where the analysis is only as good as the model reading the bands.

Why the models have to be fine-tuned

Here is the lesson that shaped the whole demonstrator: generalist AI models are narrators, not sensors. Hand a general-purpose vision-language model a SAR chip and it will confidently invent a vessel count. It is excellent at turning already-computed signals and a scene into a readable analyst brief — and useless as the detector itself.

Operational detection and classification need models fine-tuned on maritime data: labelled vessels across SAR, optical and multispectral bands; the specific classes a mission cares about; the sea states, glint and speckle that fool a naive model; and the spoofing and gap patterns of the shadow fleet. Fine-tuning is what moves a model from “plausible” to “dependable,” and it compounds — a model tuned on a client’s own vessels, waters and doctrine outperforms any generic classifier on the questions that client actually asks. We keep the generalist model in its lane (narrating the evidence, with citations) and put fine-tuned, purpose-built models on the sensing. The companion piece, fine-tuning vision models for multispectral maritime analysis, goes deeper on how.

The satellite wave arriving through 2026–2027

The reason this scale-up gets cheaper and sharper every quarter is the sensor build-out overhead.

• SAR is scaling fastest. As of mid-2026 ICEYE has launched more than 70 satellites — its four-satellite Transporter-17 batch flew on 7 July 2026 — with the Gen4 platform delivering up to 16 cm resolution and a ~400 km high-resolution field of regard, and a stated target of roughly 100 new satellites a year by 2027. Umbra images down to ~16–25 cm; Capella runs its own SAR fleet. More satellites means higher revisit, which means the multi-day windows where dark activity currently hides keep shrinking toward hours.
• Hyperspectral is arriving commercially. Planet Tanager and Pixxel put material-level sensing into the commercial catalogue for the first time at scale.

Higher revisit plus finer resolution plus spectral depth is a step-change in what corroboration is possible — and the fusion architecture we run is built to absorb each new sensor as another feed the agent can call, not a re-platforming exercise.

What stays the same: fusion, agents, explainability

Two things do not change as you climb from open to proprietary. First, value comes from fusion, not any single feed — cooperative AIS plus non-cooperative SAR plus multispectral plus registries, reasoned across by a tool-using agent at investigation time. Second, every output has to be explainable — a score a regulator, underwriter or watch officer can defend, with its data lineage attached. The live demo already shows the scored, sourced version of this on open data. Licensed feeds and fine-tuned models raise the confidence; they do not change the discipline.

The honest limits

More sensors do not make the picture complete. Tasking and revisit still bound when you can corroborate; even 16 cm SAR has classification limits; hyperspectral is cloud- and cadence-limited; registries are incomplete and sometimes deliberately obscured; and a fine-tuned model is only as good as its labels and as current as its last training run. A serious system surfaces these limits — “no SAR pass in the gap window,” “low confidence at this sea state” — rather than papering over them. That is the same honesty the open demonstrator is built on, carried into the operational build.

---

See the open-data version working at live.marineaware.com. When you are ready to run it on licensed AIS and SAR, multispectral analysis and models fine-tuned to your mission, talk to us — that is the work we do.