Fine-tuning vision models for multispectral maritime analysis

The single most important design decision in maritime AI is knowing what your models are for. Get it wrong and you build a confident liar. Get it right and generalist models and fine-tuned models each do the job they are good at. This is a short, practical piece on that division of labour — and on why multispectral sensing and domain fine-tuning are where the next gains in vessel detection come from.

Generalist models are narrators, not sensors

Hand a general-purpose vision-language model a Sentinel-1 SAR chip and ask “how many ships are here?” and it will answer — fluently, and often wrong. These models are trained on everyday photographs and prose. They have no reliable prior for radar speckle, for the way a wake registers, or for the difference between a trawler and a patrol craft at 10 metres per pixel. What they are genuinely excellent at is narration: taking signals that have already been computed — a detection count, an AIS gap, a weather state — plus a scene, and turning them into a readable, cited analyst brief.

So the rule we build to, and the one behind our live open-data dashboard, is blunt: let the generalist model narrate; never let it sense. Numbers come from purpose-built detectors on real signals. Prose comes from the language model. Cross that line and you get plausible fiction.

The maritime domain gap

Detection and classification — the sensing — need models fine-tuned on maritime data, because the domain is genuinely out-of-distribution for anything trained on consumer imagery:

• Sensors are exotic. SAR is coherent radar with speckle and layover; multispectral and hyperspectral data are cubes of dozens to hundreds of bands, not three. A model has to be taught to read them.
• The classes are specific. “Boat” is useless. Missions care about tanker versus bulker, fishing versus patrol, laden versus ballast, and the shadow-fleet behaviours — AIS gaps, spoofing, ship-to-ship transfers — that betray intent.
• The failure modes are adversarial. Sun glint, heavy seas, and deliberate spoofing all exist to fool a naive model. Robustness to them has to be trained in.

Fine-tuning closes this gap. A model adapted on labelled maritime imagery for the mission’s classes, calibrated against known cases and validated on held-out data, is the difference between “plausible” and “dependable.”

What multispectral and hyperspectral add

RGB shows a scene; SAR shows shape and presence. Neither reads material. Multispectral sensors sample several bands — typically adding near-infrared and shortwave-infrared — and hyperspectral sensors sample hundreds of contiguous narrow bands, so the data carries chemistry.

For maritime intelligence that means signatures no outline can provide: oil sheens and bilge discharge on the sea surface, gas-flaring intensity as an export proxy, chlorophyll and turbidity that shape where fishing happens, and spectral separation that sharpens vessel and cargo classification. Planet’s Tanager hyperspectral satellites resolve 426 bands across 380–2500 nm; Pixxel is building a commercial hyperspectral constellation. The catch: these bands are only as useful as the model trained to interpret them — which is precisely why multispectral and fine-tuning are one story, not two.

Fusion across sensors beats any single model

No single sensor or model wins alone. The strongest maritime picture comes from fusing them, each doing what it is best at:

1. SAR — all-weather, day-night detection and re-observation (find the silent vessel).
2. Optical + multispectral — daylight confirmation and material/chemical characterization (say what it is and what it is doing).
3. AIS + registries — identity, ownership and behaviour history (say who is behind it).

A fine-tuned detector on each sensor, orchestrated by a tool-using agent and narrated by a language model, produces one explained conclusion — not five disconnected exports.

The 2026–2027 sensor tailwind

Domain fine-tuning gets more valuable exactly as the sensor supply explodes. As of mid-2026 ICEYE has launched more than 70 SAR satellites, with Gen4 reaching up to 16 cm resolution and a target near 100 new satellites a year by 2027; Umbra images to ~16–25 cm; Capella operates its own SAR fleet; and hyperspectral goes commercial via Tanager and Pixxel. More revisit and more bands mean more raw signal — and more raw signal rewards better-tuned models, because the bottleneck shifts from can we see it to can we interpret it. The scale-up story for the whole picture is in from open source to operational.

Limits and honesty

Fine-tuned models are not oracles. They are only as good as their labels, only as current as their last training run, and only as trustworthy as their calibration. Resolution still bounds classification; clouds and cadence still bound optical and hyperspectral; and adversaries adapt. A serious deployment reports confidence and abstains when the data is thin — “low confidence at this sea state,” “no multispectral pass available” — because a defensible I don’t know beats a confident error every time. That discipline, not the model size, is what makes maritime AI usable.

---

See the narrator-not-sensor principle running on open data at live.marineaware.com. To put fine-tuned, multispectral-aware models on your own feeds and mission, talk to us.