Sensors
- Maxar WorldView-3: 31 cm panchromatic resolution at nadir, 1.24 m multispectral. At 31 cm, a standard passenger car (roughly 4.5 m long, 1.8 m wide) subtends approximately 14 by 6 pixels in pan, sufficient for individual vehicle delineation. Revisit at a given point averages 1–4.5 days depending on latitude and tasking priority.
- Airbus Pléiades Neo: 30 cm native panchromatic resolution, 1.2 m multispectral, with a daily revisit capacity across its two-satellite constellation. Colour imagery aids vehicle-type discrimination (car versus truck or SUV) through roof-colour and shape cues.
- Planet SkySat: 50 cm panchromatic, 1 m multispectral. At this resolution individual cars are marginally detectable but reliable per-vehicle counts are harder; SkySat is more useful for confirming vessel presence and estimating gross deck coverage than for precise counts. Revisit can be tasked same-day.
- AIS vessel tracking: Automatic Identification System broadcasts provide vessel identity (MMSI), position, speed, heading and declared destination. Pairing a deck-count image with the AIS record ties a load estimate to a named vessel, its departure time and its stated port of call, converting a pixel count into a trade-flow data point.
What the upper deck reveals
Pure car and truck carriers are purpose-built to stack vehicles across multiple enclosed decks, but the uppermost weather deck is open to the sky and fully visible in nadir imagery. On a large PCTC, the weather deck alone can hold several hundred passenger cars. At 30–31 cm resolution, each vehicle occupies enough pixels that a trained object-detection model, or a careful human analyst, can count them individually and distinguish cars from light commercial vehicles by roof footprint.
The commercial value of that count is straightforward. Automotive OEMs, commodity traders and port economists want to know export volumes before official statistics are published, which can lag by weeks or months. A vessel photographed at Zeebrugge or Yokohama with a full upper deck, matched to an AIS record showing a destination in the Americas or Southeast Asia, is a direct observation of trade in transit. Aggregated across a fleet and a season, it becomes an index.
The occlusion problem on multi-deck vessels
The honest constraint is that a nadir image sees only the top deck. A large PCTC such as those in the Höegh or Wallenius Wilhelmsen fleets may carry 6,000 to 8,000 vehicles across twelve or more enclosed decks. The weather deck might hold 300 to 600 cars depending on vessel class. That is, at best, roughly 5 to 10 per cent of total capacity visible from above.
This means a satellite count is not a full load figure. It is a partial-load indicator. A full upper deck is a strong signal that the vessel is heavily loaded overall, because carriers typically fill lower decks first and use the weather deck last. An empty or sparse upper deck is ambiguous: the vessel may be lightly loaded, or the weather deck may have been left clear intentionally to protect vehicles from salt spray on a particular route. Analysts should treat the upper-deck count as a lower-bound proxy, not a manifest.
There is no remote-sensing method currently capable of seeing through the hull to count vehicles on enclosed decks. SAR can detect a vessel and estimate its draught indirectly, but that is a different analytic covered elsewhere in this library.
Detection model performance at 30 cm
Object detection on very-high-resolution imagery is a well-studied problem. Published work on car detection in VHR optical imagery, using convolutional neural network architectures, consistently reports precision and recall figures above 85 per cent on clearly visible vehicles in open parking areas. PCTC decks are a harder case: vehicles are packed with minimal inter-vehicle gaps (often less than 50 cm), the deck surface is metallic and reflective, and shadows from superstructure elements occlude parts of the deck.
In practice, tightly packed rows reduce individual vehicle separation below one pixel at 30 cm, causing adjacent cars to merge into a single detection blob. Row-counting methods, which count vehicle rows rather than individual units and multiply by an estimated row capacity, can recover accuracy in dense-pack conditions. Accuracy degrades further when: the vessel is partially in shadow from a crane or terminal structure; the image acquisition angle deviates more than a few degrees from nadir, causing vehicles to lean and overlap; or cloud cover forces reliance on a slightly off-axis collect. A realistic working accuracy for upper-deck counts under good conditions is plus or minus 10 to 15 per cent of true count.
Fusing image counts with AIS to proxy trade flows
A vehicle count on its own is a curiosity. Paired with AIS data it becomes actionable. The fusion workflow is: task or retrieve an image of a PCTC at a known export terminal; run vehicle detection on the weather deck; look up the vessel's MMSI in the AIS record to confirm identity, departure time and declared destination; assign the count to an origin-destination pair.
Repeated over multiple vessels and voyages, this produces a time series of estimated vehicle exports by route. The signal is noisiest on routes where vessels call at multiple ports before final discharge, because the upper deck may be partially unloaded mid-voyage. Direct-service routes, such as Japan to Europe or Korea to North America, are the cleanest cases. AIS destination fields are self-reported and sometimes vague or deliberately set to a waypoint rather than a discharge port, so destination inference may require cross-referencing vessel schedules or historical port-call patterns.
Satellize incorporates this fusion approach in its analytics work, drawing on the same methodology used in its Tonga crop-estimation programme: combine open or licensed data streams, apply a detection or estimation model, and deliver a structured output rather than raw imagery.
Cloud cover, tasking windows and the timing problem
The most perishable moment for this analytic is the loading window. A PCTC typically spends 24 to 72 hours in port during loading operations. Miss that window with cloud cover or a tasking gap and the vessel departs with no count recorded. Coastal automotive export terminals are often in regions with significant cloud frequency: Zeebrugge averages more than 150 overcast days per year, and Nagoya and Ulsan face seasonal cloud and haze.
Mitigation options are limited but real. Multi-constellation tasking, combining WorldView-3 and Pléiades Neo attempts on the same vessel call, increases the probability of at least one clear collect. AIS can alert analysts when a known PCTC enters a port, triggering an immediate tasking request rather than relying on scheduled revisit. Even so, a realistic annual capture rate for any given vessel at a cloudy port is well below 100 per cent, and analysts building trade-flow indices must account for missing observations explicitly rather than treating gaps as zero-load events.
What a credible output looks like
The deliverable from this analytic is not a single number. It is a structured record: vessel name, MMSI, port, image timestamp, estimated upper-deck vehicle count, count confidence class (high, medium, low based on image conditions), declared AIS destination, and a flag for whether the upper deck appeared full, partial or sparse. Aggregated monthly across a defined fleet or route, this becomes a trade-flow index with a known and documented methodology.
Users should be clear about what the index measures. It measures upper-deck occupancy on imaged vessels, on the days when imagery was successfully acquired. It is not a census of all PCTC movements. Its value lies in timeliness, not completeness: it can indicate directional shifts in automotive export volumes weeks before official trade data is published, which is precisely the gap that commodity analysts and supply-chain planners need to fill.
Typical figures
| Spatial resolution (panchromatic) | 30–31 cm (WorldView-3, Pléiades Neo); 50 cm (SkySat) |
| Minimum detectable vehicle | Passenger car (approx. 4.5 m × 1.8 m) reliably detectable at 30 cm; marginal at 50 cm in dense-pack conditions |
| Count accuracy (upper deck, good conditions) | ±10–15% of true count; degrades with off-nadir angle >5°, shadow occlusion, or vehicle spacing <1 pixel |
| Deck visibility | Weather (top) deck only; enclosed lower decks not accessible to optical sensors |
| Tasking revisit | 1–4.5 days (WorldView-3); daily capacity (Pléiades Neo); same-day possible (SkySat) |
| AIS data latency | Near-real-time via terrestrial receivers; 30–90 min latency gaps in open ocean via satellite AIS |
| Cloud sensitivity | Optical only; 100% data loss under cloud. Coastal automotive ports in Northern Europe and Northeast Asia have high cloud frequency |
| Archive depth | WorldView-3 archive from 2014; Pléiades Neo from 2021; historical AIS records available from commercial providers from approx. 2009 |
| Spectral bands used | Panchromatic for counting; multispectral (visible RGB, near-infrared) for vehicle-type discrimination by colour and reflectance |
| Delivery format | Structured CSV or JSON record per vessel event; optional GeoTIFF with detection overlay; periodic trade-flow index report |
Analytics Satellize can run
| Upper-deck vehicle count per vessel call | CNN-based object detection on VHR panchromatic imagery; row-counting fallback for dense-pack conditions | Structured record (vessel, port, timestamp, count, confidence class) delivered as JSON or CSV within 24 hours of image acquisition |
| Load-state classification (full / partial / sparse) | Deck-coverage ratio derived from detected vehicle area versus total visible deck area; thresholds calibrated per vessel class | Flag appended to per-vessel record; usable as a binary signal in trade-flow screening |
| Origin-destination trade-flow index | Fusion of upper-deck counts with AIS destination data; aggregated monthly by route pair; missing-observation weighting applied | Monthly index time series in CSV; methodology note documenting capture rate and confidence intervals |
| Fleet-level export volume estimate | Extrapolation from sampled upper-deck counts using vessel-class capacity ratios; uncertainty bounds stated explicitly | Quarterly summary report by OEM export terminal or flag state, with stated assumptions on lower-deck fill rates |
| Tasking alert on PCTC port entry | AIS geofence trigger on known automotive export terminals; automatic tasking request issued to constellation operator | Alert notification with vessel identity and estimated loading-window end time; tasking confirmation receipt |
| Image-quality and occlusion assessment | Automated pre-processing: off-nadir angle check, shadow-fraction measurement, cloud-mask application; collect graded before count is run | Quality flag (A/B/C) attached to every collect; C-grade collects excluded from index calculations with gap noted |
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.