Land border crossing queue length estimation from satellite
High-resolution optical satellites can count stationary vehicles in border-crossing approach lanes where ground sensors are absent or politically inconvenient, but the revisit problem is real and must be managed honestly.
Sensors
- Maxar WorldView-3: 30 cm panchromatic resolution, 1.24 m multispectral. At this scale individual car roofs are resolvable and heavy goods vehicles are unambiguous. Single-pass revisit at a given point is roughly 1 day at mid-latitudes under tasking, but cloud and tasking contention reduce practical frequency.
- Airbus Pléiades Neo: 30 cm panchromatic, 1.2 m multispectral, four-satellite constellation. Stereo and tri-stereo collection in a single pass enables shadow-based height discrimination, which helps separate parked trucks from passenger cars without relying on roof colour alone.
- Planet SkySat: 50 cm panchromatic, 1 m multispectral. Constellation of 21 satellites allows multiple passes per day over a fixed point under tasking, partially alleviating the temporal sampling problem at higher cost per km². Video mode (up to 90 seconds) can capture queue movement directly.
- Planet Dove (PlanetScope): 3 m multispectral, near-daily global coverage. Insufficient resolution to count individual passenger cars reliably, but useful for detecting gross queue presence or absence, and for building a long baseline archive against which commercial tasking is targeted.
What a satellite actually sees at a border post
A land border crossing from 500 km looks like a narrow strip of tarmac interrupted by infrastructure: booths, canopies, barriers, and lanes that funnel vehicles into orderly rows. At 30 cm resolution, a WorldView-3 or Pléiades Neo image resolves the roof of a passenger car as a rectangle roughly 5 by 10 pixels. A heavy goods vehicle occupies something closer to 8 by 40 pixels. Those are not small targets. The challenge is not detection sensitivity; it is geometric ambiguity and temporal sampling.
Lane markings are the analyst's friend. Because approach lanes are fixed and narrow, a vehicle sitting in lane three cannot be in lane two. That spatial constraint lets automated detectors assign each identified roof to a specific queue and compute queue length in vehicle-lengths rather than raw metres. Shadow geometry adds a second constraint: at known solar elevation and azimuth, the shadow cast by a truck of standard height falls at a predictable length, which helps distinguish a truck from a low-profile car when the roof signature alone is marginal.
Where ground sensors are absent and why that matters
Many of the world's highest-friction crossings sit on borders where bilateral data-sharing is minimal or non-existent. The Kazakh-Chinese crossings at Khorgos and Dostyk, the Georgian Military Highway crossing at Zemo Larsi, and several Ukraine-Poland crossings that saw extraordinary truck queues in 2022 and 2023 are all examples where one or both sides publish no real-time queue data. Trucking associations and logistics operators have historically relied on driver reports passed through messaging groups, which are anecdotal and lag reality by hours.
Satellite imagery fills that gap precisely because it requires no cooperation from either border authority. A tasked collect over Medyka or Shehyni produces a count of vehicles in queue that is independent of any government's willingness to publish. Published research using Sentinel-2 and commercial imagery has demonstrated queue detection at Ukrainian border crossings during the 2022 disruption period, with queue lengths extending several kilometres visible even at 10 m resolution when vehicles are densely packed. At 30 cm, individual vehicle counts become feasible.
The temporal sampling problem, stated plainly
This is the central honest limitation of the method. Border queues are not static phenomena. At a busy crossing, queue length can double in 20 minutes when a shift change occurs at the booths, and collapse equally fast when a backlog clears. A single daily image captures one instant. If that instant happens to be mid-morning when processing is fast, the image will show a short queue even if the crossing was gridlocked at 06:00 and will be again at 16:00.
Planet SkySat's multi-pass tasking partially addresses this. With the 21-satellite constellation, two or three collects per day over a single point are achievable under tasking contracts, giving a morning, midday and late-afternoon snapshot. That is enough to characterise the diurnal shape of a queue if conditions repeat across days, but it is not real-time monitoring. For true intra-day dynamics, satellite imagery is a complement to other signals, not a replacement. Combining satellite counts with publicly available wait-time posts from trucking forums or mobile-phone dwell-time data from commercial location providers gives a richer picture than either source alone.
Cloud cover compounds the problem at crossings in humid climates or during winter. The Poland-Ukraine corridor is cloudy for substantial parts of the year. A tasking campaign must budget for cloud-contaminated collects and plan revisit frequency accordingly.
Detection method: from pixels to vehicle counts
The published detection pipeline for stationary vehicle counting typically involves three stages. First, a semantic segmentation model trained on labelled overhead imagery separates road surface, lane markings, vehicles and infrastructure. Models trained on datasets such as the DOTA (Detection in Optical Remote Sensing Images) benchmark have demonstrated mean average precision above 0.7 for vehicle detection at sub-metre resolution, though performance degrades at crossing canopies where vehicles are partially occluded.
Second, detected bounding boxes are projected into a lane-coordinate system using the known geometry of the crossing, derived from a reference image collected under empty conditions or from cadastral data. Each vehicle is assigned to a lane and a distance from the booth. Third, queue length is computed as the distance from the last detected vehicle to the first booth, expressed in metres or vehicle-count equivalents. Uncertainty estimates should account for occlusion under canopies, vehicles partially outside the image frame, and the possibility that some detected objects are not vehicles (parked equipment, containers). Honest reporting gives a count with an explicit confidence interval, not a single number.
Published applications and what they found
Academic and NGO work on this problem has grown since 2022. Studies published in journals indexed by MDPI Remote Sensing have used Sentinel-2 time series to identify queue presence at Ukrainian western crossings, showing that queue signatures at Medyka and Korczowa were detectable as anomalous linear features even at 10 m resolution during peak disruption. The same work noted that 10 m imagery could not reliably distinguish truck queues from passenger-car queues, which matters for freight-specific analysis.
Commercial applications at Central Asian crossings, particularly Khorgos on the Kazakhstan-China border, have used WorldView imagery to estimate throughput as a proxy for trade volume, on the logic that queue length inversely correlates with processing speed. This is a reasonable proxy but carries a confound: a long queue can mean high demand and fast processing, or moderate demand and slow processing. Without a second observation a few hours later, the two cases are indistinguishable from a single image.
Satellize's analytics work, which includes the crop-estimation programme for the Kingdom of Tonga, applies similar object-detection pipelines to optical imagery. The border-queue method sits within the same operational stack.
What this analysis can and cannot promise a buyer
A well-executed satellite queue-length programme gives a buyer a consistent, politically neutral count at crossings that publish no official data. Collected over months, it builds an archive that reveals seasonal patterns, the effect of policy changes (new lane openings, fee changes, inspection-regime shifts) and anomalies worth investigating. That archive has genuine value for logistics planning, customs forecasting and policy analysis.
What it cannot give is real-time situational awareness. A government or logistics operator who needs to know the queue length right now, to route a driver around a blockage, needs ground sensors or crowdsourced data. Satellite imagery answers the question of what the queue looked like this morning, or what the typical Tuesday queue looks like in February. Framing the product honestly in those terms is not a weakness; it is the basis for a programme that delivers on what it promises. Buyers who understand the revisit constraint can design their operations around it. Those who are sold real-time capability they will not receive are simply disappointed customers.
Typical figures
| Best available spatial resolution | 30 cm panchromatic (WorldView-3, Pléiades Neo) |
| Minimum detectable vehicle | Passenger car reliably detectable at ≤50 cm resolution; marginal at 1 m; not individually countable at 3 m or coarser |
| Revisit under tasking (commercial) | 1 pass/day typical (WorldView-3, Pléiades Neo); up to 2–3 passes/day with Planet SkySat multi-pass tasking |
| Spectral bands used | Panchromatic for detection; VNIR multispectral (4–8 bands, 400–900 nm) for vehicle-type discrimination |
| Image latency (tasked collect to delivery) | Typically 2–24 hours from collect to orthorectified image delivery depending on provider pipeline |
| Queue length measurement precision | ±1–2 vehicle lengths at 30 cm resolution in open lanes; degrades under canopy or shadow occlusion |
| Cloud cover impact | Optical only; cloud-contaminated collects yield no usable data. Temperate and sub-tropical crossings may lose 30–50% of planned collects seasonally |
| Archive depth | Commercial providers hold imagery archives from approximately 2014 (WorldView-3) and 2021 (Pléiades Neo) onward; Planet SkySat from 2016 |
| Area per collect | WorldView-3 swath 13.1 km; Pléiades Neo 14 km; SkySat 6.6 km. A single crossing fits comfortably within one strip |
| Delivery format | GeoTIFF orthoimage; vector GIS layer of detected vehicle bounding boxes; CSV queue-length time series |
Analytics Satellize can run
| Vehicle count per lane | Convolutional neural network object detection (DOTA-class benchmark), bounding-box assignment to lane-coordinate system derived from reference imagery | GIS vector layer with per-vehicle bounding boxes and lane attribution, delivered per collect |
| Queue length in metres and vehicle equivalents | Lane-coordinate projection of furthest detected vehicle from booth position; uncertainty interval reported from occlusion model | Tabular report per collect with point estimate and ±confidence interval |
| Vehicle-type split (HGV vs passenger) | Roof-area and shadow-length classification using Pléiades Neo stereo-derived height, supplemented by aspect-ratio filtering | Per-collect breakdown of heavy goods vehicles vs passenger cars per lane |
| Diurnal queue profile (multi-pass days) | Time-series aggregation of 2–3 SkySat collects per day; cubic-spline interpolation between observations with explicit gap flagging | Daily chart of estimated queue length at observed times, with interpolated trend and confidence shading |
| Seasonal baseline and anomaly detection | Historical archive analysis using PlanetScope 3 m imagery for queue-presence classification; commercial imagery for count validation at anomaly dates | Monthly baseline report with flagged anomaly dates and supporting imagery thumbnails |
| Policy-change impact assessment | Before-and-after comparison of queue distributions across a defined event window (lane opening, fee change, inspection-regime shift); Mann-Whitney test for distributional shift | Analytical memo with statistical test results and time-series chart |
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.