Optical vessel classification by hull type
Very-high-resolution optical satellites can distinguish tankers from bulk carriers, container ships, fishing vessels, and naval craft by hull geometry and deck profile. Classification accuracy depends critically on achieving sub-1-metre ground sampling distance and clear skies.
Sensors
- Maxar WorldView-3: 0.31 m panchromatic GSD, 8-band multispectral at 1.24 m. The sharpest commercially available optical imagery for resolving deck equipment such as crane booms, hatch covers, and anchor chains. Revisit roughly 1 day at mid-latitudes with off-nadir tasking.
- Airbus Pléiades Neo: 0.30 m panchromatic GSD, 4-band multispectral at 1.2 m. Stereo and tri-stereo collection in a single pass allows 3-D superstructure profiling. Constellation of two satellites gives daily revisit over most ports.
- Planet SkySat: 0.50 m panchromatic GSD, 4-band multispectral at 1.0 m. A constellation of 21 satellites enables rapid re-tasking and same-day collect windows. Resolution is sufficient for hull-type classification but marginal for fine equipment identification.
- Airbus SPOT-7: 1.5 m panchromatic GSD, 6 m multispectral. Below the threshold for reliable classification of individual vessels under roughly 100 m LOA; useful for large tankers and bulk carriers in open anchorages, or for wide-area scene context around a port.
What a hull silhouette actually reveals
Every vessel class leaves a characteristic geometric signature when viewed from directly overhead. A crude oil tanker presents a long, unbroken flush deck with a single superstructure block aft and no visible hatch coamings. A bulk carrier shows a row of raised hatch covers running the full cargo length, typically seven to nine rectangular panels. A container ship's deck is occupied by stacked box arrays whose shadow height varies with load. A fishing trawler has a compact hull, a prominent gantry or net drum aft, and a length-to-beam ratio that differs markedly from commercial cargo types.
Naval vessels are the hardest class to confuse with commercial shipping: flight deck geometry, gun mount positions, and the absence of cargo handling equipment are all visible at sub-1-metre resolution. The challenge is not distinguishing a frigate from a tanker. It is distinguishing a refrigerated cargo vessel from a small fishery support ship, or a product tanker from a small bulk carrier when both are in ballast and riding high.
Why resolution is not a preference but a physical constraint
The International Maritime Organization defines vessel length categories that set a hard floor on what imagery can resolve. A 180-metre Panamax bulk carrier occupies roughly 580 pixels along its centreline in a WorldView-3 panchromatic image. The same vessel at SPOT-7 resolution occupies 120 pixels. Hatch coamings, which are the primary discriminating feature for bulk carriers, are typically 0.4 to 0.8 metres tall and cast shadows of similar width. At 1.5 m GSD those shadows merge into background noise. At 0.31 m GSD they are unambiguous.
The practical classification threshold sits at approximately 0.5 m GSD for vessels above 100 m LOA, and closer to 0.3 m for smaller craft such as fishing vessels and patrol boats. Below those thresholds, convolutional neural networks trained on labelled ship imagery can still detect vessel presence and estimate rough length, but hull-type labels carry meaningful uncertainty. Published work on the DOTA and HRSC benchmark datasets, which use imagery in the 0.1 to 1.0 m range, consistently shows classification accuracy dropping from above 90 percent to below 70 percent as GSD crosses 1 metre.
How CNN classifiers are trained and where they fail
Convolutional neural networks for ship classification are trained on labelled patches cropped from georeferenced satellite imagery, with labels drawn from AIS vessel type codes, Lloyd's register data, or manual annotation. The network learns spatial features: aspect ratio, the presence and position of superstructure blocks, the texture of hatch covers versus container stacks, the outline of a crane boom. Models trained on WorldView imagery do not transfer cleanly to SkySat imagery without fine-tuning, because the spatial frequency content differs even when nominal resolution is similar.
Failure modes are predictable. Vessels viewed at high off-nadir angles (beyond roughly 25 degrees) show foreshortened decks and exaggerated superstructure height, which degrades aspect-ratio features. Vessels in ballast sit higher in the water and expose more freeboard, altering the hull silhouette relative to the same ship fully laden. Partial occlusion by other vessels in a crowded anchorage causes misclassification. And a vessel that has been converted from one type to another, a common occurrence in older tonnage, may carry the hull geometry of its original class while operating as something else entirely.
Pansharpening, the process of fusing panchromatic and multispectral bands to produce a colour image at panchromatic resolution, introduces its own artefacts around high-contrast edges such as hull waterlines and crane booms. Classifiers trained on raw panchromatic data sometimes perform better than those trained on pansharpened composites because the fusion artefacts add noise to the features the network relies on most.
Cloud cover and sun-glint: the two conditions that simply stop the method
Optical classification has one categorical weakness that no algorithm addresses: cloud cover blocks the signal entirely. Tropical ports and fishing grounds in the inter-tropical convergence zone can experience cloud cover on more than 70 percent of daylight hours during wet seasons. In those conditions, a tasked collect returns a white scene and no classification is possible. SAR-based detection is the complement for persistent cloud environments, though SAR cannot replicate the hull-geometry detail that optical imagery provides.
Sun-glint is the second denial condition. When the solar zenith angle and satellite look angle combine to place specular reflection from the sea surface within the sensor's field of view, water pixels saturate and vessels near the glint zone lose contrast against the background. Glint geometry is predictable from ephemeris data and can be avoided by scheduling collects outside the glint window, but this constrains the available collect times and may conflict with the revisit requirement. Atmospheric haze over industrial ports adds a third, softer constraint: it reduces effective contrast without eliminating the image, but it degrades the fine-detail features that classification relies on.
Building a classification pipeline that acknowledges its own uncertainty
A production classification pipeline combines several steps. Automated ship detection, typically using a sliding-window or anchor-based object detector, identifies candidate vessel chips in the full scene. Each chip then passes through a multi-class CNN that outputs a probability distribution across vessel types rather than a hard label. Chips where the top-class probability falls below a threshold, often set at 0.75 to 0.85 in published maritime classification work, are flagged for human review rather than auto-labelled.
Cross-referencing against AIS records, port state databases, and vessel registries adds a second evidence layer. A chip classified as a bulk carrier that AIS records show as a registered chemical tanker warrants scrutiny, not automatic correction. The disagreement is itself intelligence. Satellize applies this kind of multi-source cross-referencing in its analytics work, where a single sensor output is treated as a hypothesis rather than a conclusion.
Archive depth matters for this use case. WorldView imagery archives extend back to 2009 for some regions, Pléiades to 2012. Comparing current hull configuration against historical imagery of the same vessel at the same berth can reveal structural modifications, which is relevant for sanctions monitoring and naval intelligence. The archive is not comprehensive for every port, and cloud-affected collects reduce effective temporal density in tropical regions.
Typical figures
| Best available panchromatic GSD | 0.30 m (Pléiades Neo, WorldView-3) |
| Minimum GSD for hull-type classification | ~0.5 m for vessels >100 m LOA; ~0.3 m for smaller craft |
| Multispectral resolution (commercial VHR) | 1.0–1.24 m (WorldView-3, Pléiades Neo) |
| Typical tasked revisit (VHR commercial) | 1 day at mid-latitudes with off-nadir; 3–5 days near-nadir |
| Collect latency (tasked to delivered) | 2–6 hours for priority tasking; 24–48 hours standard |
| Spectral bands used for classification | Panchromatic primary; 4- or 8-band multispectral for material discrimination |
| Minimum detectable vessel length | ~10 m at 0.3 m GSD; ~50 m at 1.5 m GSD |
| Primary data-denial conditions | Cloud cover (categorical); sun-glint (geometric, schedulable around); haze (partial) |
| Archive depth (major commercial providers) | WorldView from 2009; Pléiades from 2012; SkySat from 2016 |
| Delivery formats | GeoTIFF ortho imagery; vector ship polygons (GeoJSON/Shapefile); classification report (PDF/JSON) |
Analytics Satellize can run
| Hull-type classification label with confidence score | Convolutional neural network multi-class classifier trained on labelled VHR ship imagery (DOTA/HRSC benchmark architecture class) | GeoJSON feature layer with vessel centroid, bounding polygon, class label, and per-class probability vector |
| Vessel length and beam estimation | Oriented bounding-box regression on detected ship chips; calibrated against known vessel dimensions in training set | Attribute table appended to detection layer; CSV export for fleet-size analysis |
| Unregistered or misclassified vessel flag | Cross-reference of optical classification label against AIS vessel-type code and Lloyd's/IHS registry; disagreement scoring | Alert report listing vessels where optical class and registry class diverge beyond defined threshold |
| Structural modification detection | Change detection between archived and current VHR collects of same vessel; feature-level diff on superstructure and deck equipment | Before/after image pair with annotated change regions; narrative summary |
| Fleet composition report for a named port or anchorage | Scene-level aggregation of per-vessel classification outputs; temporal stacking across multiple collect dates | Time-series table of vessel-type counts by date; PDF summary with annotated imagery |
| Collect feasibility and cloud-risk assessment | Climatological cloud-cover probability from ERA5/MODIS cloud fraction archives combined with glint geometry modelling | Tasking recommendation memo with optimal collect windows and expected denial probability by month |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.