Road freight activity proxied by truck counts at logistics nodes
Sub-metre optical imagery from Planet, Maxar and Pléiades Neo can count heavy goods vehicles at distribution centres and border crossings, turning parking geometry into a daily freight-flow index.
Sensors
- Planet SuperDove (PlanetScope): 3 m native resolution, 8-band multispectral. Daily global revisit makes it the workhorse for time-series counts, though 3 m is at the edge of reliable truck-versus-car discrimination; shadow analysis and object length are essential supplements.
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral. At this resolution individual axle shadows and trailer lengths are measurable. Revisit is 1–4.5 days depending on latitude and tasking priority, so it suits spot-checks and model validation rather than daily monitoring.
- Airbus Pléiades Neo: 0.30 m panchromatic, 1.2 m multispectral, four-satellite constellation giving sub-daily revisit at mid-latitudes under clear skies. Stereo tasking adds height information that helps separate stacked containers from parked vehicles.
- Planet SkySat: 0.50 m panchromatic, 1 m multispectral. Up to 12 revisits per day over a fixed site when tasked in continuous-collection mode, making intra-day shift detection feasible at specific nodes of interest.
Why truck counts work as a freight proxy
A distribution centre is a buffer. Goods arrive, dwell briefly, and depart. The number of heavy goods vehicles visible at any moment is a direct function of throughput: high dwell, high count; rapid turnover, lower count but higher churn between passes. Neither metric is available from public sources in real time. Customs declarations arrive late, toll data is national and aggregated, and operators have no obligation to publish utilisation figures.
Satellite counts sidestep all of that. A truck occupies roughly 16–20 metres of length and 2.5 metres of width. At 0.3–0.5 m resolution those dimensions are measurable directly. At 3 m resolution they are inferrable from shadow length, parking-bay geometry and the aspect ratio of the bright-dark pixel pattern a cab-plus-trailer produces. The physics is simple and does not change.
Separating trucks from everything else
Passenger cars are typically 4–5 m long and under 2 m wide. A standard European semi-trailer is 16.5 m long. That is a 3:1 length ratio, which survives resampling to 3 m pixels provided the image is taken at low solar zenith angle so shadows project clearly. Published studies using PlanetScope imagery apply a minimum bounding rectangle filter: objects with a long-axis dimension below 10 m are excluded from the truck class. At WorldView-3 resolution the cab, trailer joint and rear axle cluster are individually resolved, removing most ambiguity.
Parking geometry adds a second discriminator. Loading bays are designed for 13.6 m trailers; the bay markings are visible in sub-metre imagery and act as a reference ruler. A vehicle that fills 80–100 % of a marked bay is almost certainly a heavy goods vehicle. Vehicles parked at random angles in unmarked areas are harder to classify, and analysts typically flag those as uncertain rather than forcing a count.
Colour and spectral response help at the margins. Refrigerated trailers are often white or silver and have a distinctive near-infrared reflectance signature compared with dark tarmac. This is useful for separating parked trailers from road markings at 3 m resolution, though it is not a reliable truck-type classifier on its own.
The revisit floor and what falls below it
A single daily pass from PlanetScope captures one snapshot, typically between 09:30 and 11:30 local time depending on the orbital ground track. That is fine for a daily occupancy index but it misses the morning arrival surge, the midday lull and the afternoon departure wave entirely. Shift patterns at 24-hour fulfilment centres, which may turn over their entire yard population three times in a day, are simply unresolvable from one pass.
SkySat's multi-collect tasking changes this for specific sites. Twelve passes per day over a single location is achievable in principle, though cloud cover, orbital geometry and competing tasking requests reduce the practical figure. For a border crossing or a single large warehouse, intra-day rhythm becomes visible at four or more passes. For a network of fifty distribution centres, the cost and scheduling complexity make daily single-pass PlanetScope the realistic baseline.
Cloud is the other hard limit. Optical sensors see nothing through thick cloud. In temperate climates, monthly cloud-free pass rates at a given site can fall below 50 % in winter. Analysts working on freight indices typically gap-fill using a rolling median of the previous five cloud-free observations, which smooths short-term spikes. Anyone selling a guaranteed daily count from optical imagery alone is not being straight with you.
Change detection at scale: how the algorithm pipeline works
Manual counting is viable for a handful of sites and a validation exercise. For a network of hundreds of nodes, automated change detection is the only practical approach. The standard pipeline registers each new image to a base orthorectified reference, subtracts it, and passes the residual through an object detector trained on labelled examples of trucks in the target region. Convolutional neural network detectors trained on WorldView-3 imagery have published precision and recall figures in the 85–95 % range on held-out test sets, though performance degrades when the model encounters a new facility layout or an unfamiliar trailer type.
The count itself is less useful than the count relative to a baseline. A distribution centre with 40 bays that shows 38 trucks on a Tuesday morning is running near capacity. The same site showing 12 trucks on the same day the following week is a signal worth investigating. Change detection therefore requires a consistent historical baseline, which means the archive depth of the constellation matters as much as the current resolution. Planet's archive runs from 2016; Maxar's from 2008 for some regions.
Honest limits and what to do about them
Three failure modes recur in published work. First, occlusion: canopy roofs over loading bays, which are common at large fulfilment centres, hide trucks completely. A site with covered bays will systematically under-report. Ground operators or building-permit data can flag which sites have this problem before the analysis begins.
Second, double-counting at high-churn sites: if a truck departs and a new one arrives between two passes taken 90 minutes apart, the count looks stable when throughput was actually double. This is not detectable from imagery alone; it requires pairing the satellite index with at least one ground-truth source, such as a sample of weigh-station records or carrier API data, to calibrate the dwell-time assumption.
Third, model drift: a detector trained on North American trailer types performs poorly on European curtainsider trailers or Asian flatbeds without retraining. Regional model variants are necessary for cross-border freight monitoring programmes.
Satellize addresses the calibration problem the same way it approached crop estimation in Tonga: by building the ground-truth collection into the programme design from the start, not as an afterthought. A satellite count without a calibration layer is a relative index, not an absolute throughput figure. Both have value, but they answer different questions.
What the output actually looks like
The primary deliverable is a time-series index: truck count per site per observation, normalised to a baseline period, with a confidence interval that reflects cloud-cover gaps and detection uncertainty. Secondary outputs include a site-level heat map showing which nodes in a network are running above or below their historical norm, and an alert layer that flags sites exceeding a threshold deviation for more than three consecutive observations.
For border crossings, queue length in vehicle-equivalents is a more useful metric than raw count, because the approach road geometry is fixed and the queue tail is directly measurable. Published work on the US-Mexico border using WorldView imagery has demonstrated queue-length estimates within 10–15 % of ground counts under clear-sky conditions. That accuracy band is honest and useful for policy planning, even if it is not precise enough for real-time traffic management.
Typical figures
| Spatial resolution (best available) | 0.30 m panchromatic (WorldView-3, Pléiades Neo); 0.50 m (SkySat); 3 m (PlanetScope SuperDove) |
| Minimum detectable vehicle length | ~8 m at 0.5 m resolution; ~12 m reliably at 3 m resolution with shadow assist |
| Revisit frequency | Daily (PlanetScope); 1–4.5 days (WorldView-3, tasking-dependent); up to 12 passes/day per site (SkySat multi-collect) |
| Spectral bands | Panchromatic + 4–8 multispectral bands (visible to near-infrared, 450–860 nm typical) |
| Cloud impact | Full occlusion under thick cloud; temperate-climate clear-pass rates can fall below 50 % in winter months |
| Archive depth | Planet from 2016; Maxar from 2008 (selected regions); Pléiades from 2012 |
| Latency (tasked collection to delivery) | Typically 2–6 hours for priority tasking; next-day for standard archive pull |
| Coverage per collection | PlanetScope: global daily strips; SkySat/WorldView-3/Pléiades Neo: point-tasked scenes, typically 25–100 km² per collect |
| Delivery formats | GeoTIFF orthoimage, GeoJSON vehicle detections, CSV time-series index, dashboard feed |
Analytics Satellize can run
| Daily truck-count index per logistics node | Object detection (CNN trained on sub-metre labelled imagery) applied to registered daily PlanetScope scenes | CSV or API feed: site ID, observation timestamp, truck count, cloud-flag, confidence interval |
| Network-level occupancy heat map | Count normalised to rolling 90-day baseline; z-score deviation per site | GIS layer (GeoPackage or shapefile) updated per new cloud-free observation; web dashboard view |
| Intra-day shift-pattern profile (specific sites) | Multi-collect SkySat time series; count at each pass plotted against time-of-day | Site report showing arrival/departure curves; delivered weekly or on-demand |
| Anomaly alert: sustained deviation from baseline | Three-consecutive-observation threshold trigger on z-score; human-reviewed before dispatch | Email or webhook alert with annotated image chip and deviation magnitude |
| Border crossing queue-length estimate | Approach-road segmentation; vehicle-equivalent count along queue polygon; calibrated against published ground-count studies | Queue-length time series (vehicle-equivalents) with ±15 % uncertainty band noted explicitly |
| Covered-bay occlusion audit | Canopy detection from building footprint and shadow analysis on WorldView-3 or Pléiades Neo imagery | Site characterisation report flagging which bays are unobservable from nadir; adjusts count denominator |
| Historical baseline construction | Archive retrieval and batch processing over Planet or Maxar archive; gap-filling via rolling median | Multi-year site-level time series delivered as GeoTIFF stack and summary statistics table |
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.