Roll-on/roll-off vehicle terminal inventory from optical imagery
Very-high-resolution optical satellites resolve individual passenger cars on RoRo aprons, turning parked inventory into a quantitative signal for automotive exporters and commodity traders tracking production-to-shipment lead times.
Sensors
- Maxar WorldView-3: 0.31 m panchromatic GSD, 1.24 m multispectral; agile tasking with same-day or next-day collection over specific terminals; the primary confirmation layer for individual vehicle classification and count validation
- Airbus Pléiades Neo: 0.30 m panchromatic GSD, 1.20 m multispectral across four satellites; revisit of 1–2 days at mid-latitudes; comparable to WorldView-3 for vehicle-level resolution and useful as a second independent confirmation source
- Planet SuperDove (PlanetScope): 3 m multispectral GSD, daily global revisit; cannot resolve individual cars but tracks apron-wide reflectance change and gross occupancy trends across time series, flagging when a high-resolution task is worth ordering
- BlackSky Global: 0.5 m panchromatic GSD, with same-day revisit capability over priority sites; useful for rapid intraday snapshots ahead of known vessel departures, though archive depth is shallower than Maxar
What a full apron actually tells you
A RoRo terminal apron is, in effect, a publicly visible buffer stock. When a manufacturer's logistics pipeline is healthy, vehicles move from factory to vessel in days. When port congestion, vessel shortages or demand softness intervenes, the apron fills. The count at any given moment is a lagging signal on production and a leading signal on export shipment volumes, which makes it genuinely useful to traders pricing automotive freight, analysts forecasting OEM quarterly delivery numbers, and supply-chain teams managing vessel slot commitments.
The analytic is simpler than it sounds. A modern passenger car occupies roughly 8–10 square metres of apron space. At 0.31–0.30 m GSD, both WorldView-3 and Pléiades Neo resolve the roof, windscreen and shadow of an individual vehicle with enough fidelity that a convolutional object-detection model trained on labelled satellite imagery can distinguish a passenger car from a light commercial vehicle, a heavy truck or a piece of agricultural equipment. That class distinction matters: RoRo terminals often handle mixed cargo, and conflating a row of combine harvesters with a row of saloon cars produces a meaningless count.
The two-layer collection strategy
Running WorldView-3 or Pléiades Neo over every terminal every day is expensive and unnecessary. The practical approach uses Planet SuperDove's daily 3 m imagery as a low-cost screening layer. At 3 m, individual cars are not resolved, but the aggregate reflectance signature of a densely packed apron changes measurably as occupancy rises and falls. A normalised difference in visible-band brightness between successive days, calibrated against known empty-apron reflectance, can flag terminals where occupancy has shifted by more than roughly 15–20 percentage points. Those flags trigger a commercial tasking order for a sub-metre image.
This tiered approach compresses cost significantly and is well-suited to monitoring a portfolio of terminals across multiple geographies simultaneously. The latency penalty is small: Planet imagery is typically available within 12–24 hours of acquisition, and a WorldView-3 or Pléiades Neo task placed the same day can deliver a confirmatory image within 24–48 hours. For a trader who needs weekly inventory snapshots rather than intraday precision, that cadence is more than adequate.
Detection limits and honest failure modes
Cloud cover is the primary operational constraint. Optical imagery is useless through thick cloud, and major RoRo hubs in northern Europe, Japan and South Korea experience significant overcast periods. A monitoring programme that relies on optical alone will have data gaps of several consecutive days in winter months at high-latitude terminals. Combining optical with SAR-based apron-change detection can partially bridge those gaps, though SAR cannot classify vehicle types at current commercial resolutions.
Shadow and occlusion introduce counting errors at the margins of vehicle blocks, particularly when imagery is collected at high off-nadir angles. Published studies on parking-lot vehicle detection from sub-metre imagery report precision and recall figures typically in the 85–95 percent range under good conditions, degrading toward the lower end when vehicles are tightly packed and shadows overlap. A count derived from satellite imagery should therefore be treated as an estimate with a plausible error band of plus or minus 5–10 percent, not a census. Partial loads under terminal canopies or multi-storey vehicle-storage structures are invisible to nadir and near-nadir optical sensors entirely.
Vehicle colour variation and seasonal lighting changes also affect model performance. A detector trained predominantly on one terminal's imagery may underperform at a terminal with different apron surfacing or vehicle mix. Retraining or fine-tuning on site-specific labelled data materially improves accuracy.
Building the time series: archive depth and trend analysis
Maxar's WorldView archive extends back to WorldView-1 in 2007, with WorldView-2 from 2009 and WorldView-3 from 2014. Planet's SuperDove archive offers near-daily global coverage from around 2017. That depth is sufficient to reconstruct multi-year inventory cycles at major terminals, including the sharp apron-fill events visible during the 2021 semiconductor shortage, when finished-vehicle inventory backed up at European and Korean ports for weeks.
A time series of vehicle counts, plotted against publicly available vessel-departure schedules from AIS data and against OEM production announcements, produces a composite lead-time indicator. The lag between apron fill and vessel departure is measurable from the imagery itself: count on the day before a known PCTC departure, then count again 48 hours after. The residual tells you how full the vessel left. That is a data point most analysts currently have no systematic way to obtain.
Where this sits in a broader analytics workflow
Vehicle inventory on the apron is one node in a chain. Satellize builds the apron-count layer as a structured time-series feed that clients can join to vessel-tracking data, port-call records and OEM production reports. The crop-estimation programme Satellize runs for the Kingdom of Tonga uses a structurally similar pipeline, where object-level detection from high-resolution imagery feeds into an aggregated national indicator. The underlying pattern recognition and time-series architecture transfers directly to vehicle counting, with the object class and the economic interpretation being the main differences.
For automotive manufacturers, the most immediate application is benchmarking their own terminal inventory against competitor terminals visible in the same imagery. For commodity traders and freight brokers, the count provides an independent check on whether OEM shipping guidance is consistent with what is physically sitting on the apron. Neither use case requires the satellite operator to have any privileged access to the terminal. The aprons are open to the sky.
Typical figures
| Spatial resolution (confirmation layer) | 0.30–0.31 m GSD (Pléiades Neo / WorldView-3 panchromatic) |
| Spatial resolution (screening layer) | 3 m GSD (Planet SuperDove multispectral) |
| Revisit (screening) | Daily global coverage (Planet SuperDove) |
| Revisit (confirmation) | 1–2 days typical at mid-latitudes; same-day possible with tasking priority (WorldView-3, Pléiades Neo) |
| Delivery latency | 12–48 hours from acquisition to processed count, depending on tasking tier |
| Spectral bands used | Panchromatic (vehicle delineation); red, green, blue, NIR (surface classification and shadow masking) |
| Minimum detectable target | Individual passenger car (~4.5 m length) at 0.31 m GSD; not reliably resolved at 3 m GSD |
| Count accuracy (published range) | 85–95% precision and recall under good illumination and low cloud; degrades with shadow overlap and occlusion |
| Archive depth | WorldView-3 from 2014; Planet near-daily from ~2017; WorldView-2 from 2009 for historical reconstruction |
| Delivery formats | GeoJSON vehicle polygons, CSV count time series, GeoTIFF confidence maps, structured API feed |
Analytics Satellize can run
| Terminal-level vehicle count | Convolutional object detection (e.g. Faster R-CNN or YOLO family) trained on labelled sub-metre satellite imagery of vehicle aprons | Per-image vehicle count by class (passenger car, light commercial, heavy equipment) as GeoJSON with bounding boxes and confidence scores |
| Apron occupancy index | Visible-band reflectance normalisation against known empty-apron baseline from Planet SuperDove daily time series | Daily occupancy index (0–100%) per terminal as CSV or API feed; threshold alerts triggering commercial task orders |
| Production-to-shipment lead-time estimate | Cross-referencing apron count time series with AIS-derived vessel departure events to compute inventory drawdown rate per vessel call | Weekly lead-time indicator report per terminal, with vessel-fill estimate and trend chart |
| Multi-terminal inventory comparison | Standardised count pipeline applied across a defined portfolio of terminals; normalised by reported terminal capacity from public port documentation | Ranked dashboard of terminal occupancy across monitored sites, updated on each new image acquisition |
| Historical inventory reconstruction | Retrospective object detection applied to WorldView and Planet archive imagery; counts aggregated into monthly time series | Multi-year monthly count dataset per terminal, suitable for econometric modelling and cycle analysis |
| Anomaly alert on rapid fill or drawdown | Statistical process control on rolling occupancy index; flags deviations beyond two standard deviations from trailing 30-day mean | Automated alert with supporting image thumbnail and count delta, delivered by email or webhook within 24 hours of triggering image |
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.