Military logistics depot throughput estimation
Counting vehicles and railcars in time-series optical imagery reveals supply-chain tempo at depots, fuel farms, and railheads. The method gained public validation during open-source analysis of Russian logistics before the 2022 Ukraine invasion.
Sensors
- Planet Dove (PlanetScope): 3–5 m native pixel resolution, daily global revisit across the 200+ satellite constellation. Sufficient to distinguish vehicle presence from background clutter; not fine enough to read vehicle sub-type without shadow analysis or a spectral library.
- Planet SkySat: 0.5 m resolution, tasked on demand, up to 12 revisits per day over a fixed point. At this scale, cab and trailer geometry become readable, enabling truck-class discrimination without relying solely on shadow length.
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral. The highest commercially available optical resolution; supports vehicle-type identification against trained libraries and allows precise shadow-length measurement for height estimation of covered assets.
- Airbus Pléiades Neo: 0.3 m panchromatic, tri-stereo capable. Stereo pairs add a height dimension that helps separate stacked cargo from flat vehicles, and the 4-satellite constellation achieves same-day revisit over most latitudes.
Why counting vehicles at a depot is not as simple as it sounds
A military depot photograph is not a ledger. Vehicles present at the moment of imaging are a snapshot, not a flow rate. Turning presence into throughput requires repeated observations at known intervals, combined with assumptions about average dwell time. If a truck spends four hours at a railhead on average, and you observe 40 trucks present at a given moment, you can estimate roughly 240 truck-movements per day. The logic is borrowed from queuing theory, and the uncertainty is real: dwell-time assumptions vary by depot type, season, and operational tempo, so analysts should treat throughput figures as order-of-magnitude estimates rather than precise counts.
Cloud cover is the most obvious constraint. Optical imagery is useless through thick cloud, and many logistics hubs in contested regions sit under persistent overcast for weeks at a time. Daily Planet revisit helps by increasing the probability of at least one clear acquisition per day, but there is no optical workaround for solid cloud. SAR would be the complement, though SAR vehicle discrimination at depot scale is a separate and harder problem not covered here.
What a shadow gives away
At resolutions coarser than about 0.5 m, individual vehicle features, such as cab shape, axle count, or load type, are difficult to read directly. Shadow length fills the gap. Given the solar elevation angle at the time of acquisition (recoverable from image metadata), a shadow of known pixel length maps to a known object height. A standard 40-tonne military truck sits roughly 3.5–4 m tall when loaded; a fuel tanker trailer has a distinctive cylindrical silhouette even at 1 m resolution. Shadow-based height estimation was a documented technique in the public OSINT analysis of Russian rail-borne logistics in late 2021 and early 2022, used to distinguish covered flatcars carrying armour from empty cars.
Vehicle-type libraries, built from confirmed examples in high-resolution imagery, extend the method. A library entry pairs a vehicle's known dimensions with its spectral signature and shadow profile at multiple sun angles. Matching incoming imagery against such a library is a form of template matching, computationally straightforward but sensitive to image quality, compression artefacts, and off-nadir viewing angle. WorldView-3 and Pléiades Neo, both capable of sub-0.35 m resolution, make library matching reliable enough to distinguish, for example, a KamAZ-5350 military cargo truck from a civilian equivalent, or a TZM reload vehicle from a standard flatbed.
Temporal density is the analytic product
A single image of a depot is a curiosity. A time series is intelligence. The key metric is the change in vehicle count between successive acquisitions, plotted against time. A depot preparing to support a major operation will show a rising vehicle count over days to weeks, often with a characteristic surge-and-drain pattern as loads are consolidated and then dispatched. Railcar counts follow a different rhythm: cars arrive loaded, sit for unloading (typically 12–48 hours for a standard military supply train), and depart empty or repositioned.
Planet's daily global coverage, achieved through the Dove constellation's approximately 200 satellites in sun-synchronous orbit, is the baseline for this kind of temporal counting. It does not guarantee a cloud-free image every day, but it maximises the probability of usable acquisitions. SkySat tasking adds intra-day revisit where the situation demands it, at the cost of narrower swath and higher collection priority. The practical workflow combines Dove for continuous background monitoring with SkySat or WorldView-3 tasking when a count anomaly triggers closer inspection.
The public record: Russian logistics before February 2022
The most publicly documented application of this method is the open-source analysis of Russian military logistics conducted by organisations including Maxar, the Middlebury Institute of International Studies, and various OSINT analysts in the months before Russia's full-scale invasion of Ukraine. Satellite imagery published in late 2021 and early 2022 showed vehicle concentrations at railheads and depot facilities in Belgorod, Yelnya, and other locations. Analysts counted vehicle types, tracked the appearance of fuel and ammunition support equipment, and noted the construction of field hospitals, all from commercially licensed imagery.
That analysis was imperfect. Counts varied between analysts. Cloud obscured key facilities on some dates. Some vehicle identifications were contested. But the aggregate picture, derived from dozens of separate acquisitions over several months, was directionally correct and provided advance warning that proved accurate. The episode established a public methodological baseline that defence analysts now treat as a reference case for what commercial optical imagery can and cannot do at this task.
Honest limits before you commission an assessment
Resolution floors matter. At PlanetScope's 3–5 m resolution, a truck is a few pixels. You can count objects; you cannot classify them with confidence. SkySat and WorldView-3 at sub-0.5 m change the picture substantially, but tasking those satellites over a denied-access area requires the target to be within their collection geometry on a given pass, and collection requests over politically sensitive locations may face restrictions depending on the operator's end-user licence terms.
Throughput estimates carry compounding uncertainty. Dwell-time assumptions are the weakest link. Without signals intelligence or human reporting to calibrate them, analysts typically use published military logistics doctrine as a proxy, which may not reflect actual practice. Count accuracy degrades with image compression, high off-nadir angles beyond about 25 degrees, and camouflage netting, which is specifically designed to defeat optical observation. A well-netted depot will undercount significantly. Analysts should report confidence intervals, not point estimates, and flag any collection gaps longer than 72 hours.
Satellize structures assessments around explicit uncertainty bands and flags every gap in the time series. The same statistical discipline used in the Tonga crop-estimation programme, where yield uncertainty is reported alongside the central estimate, applies here.
From raw counts to a decision-ready product
The deliverable that matters to a defence or national-security client is not a spreadsheet of vehicle counts. It is an interpreted time series with anomaly flags, a throughput range estimate with stated assumptions, and a change narrative that explains what the pattern implies about operational tempo. That means combining the automated counting output with an analyst layer that knows the difference between a routine resupply cycle and a pre-offensive surge.
Geospatial outputs, vehicle-count layers in GeoJSON or GeoPackage format, integrate directly into existing command-and-control GIS environments. Periodic reports, structured around named facilities and dated image acquisitions, provide the audit trail that intelligence consumers require. Alert thresholds, set to trigger when vehicle counts exceed a rolling baseline by a defined percentage, allow continuous monitoring without requiring an analyst to review every image.
Typical figures
| Best available spatial resolution (optical) | 0.3 m (WorldView-3 panchromatic, Pléiades Neo panchromatic) |
| Routine monitoring resolution | 3–5 m (PlanetScope Dove); sufficient for vehicle presence counting, not type classification |
| Revisit rate (PlanetScope) | Daily global; typically 1–2 clear acquisitions per day at mid-latitudes in fair weather |
| Revisit rate (SkySat) | Up to 12 times per day over a tasked point; narrow swath (~6 km) |
| Minimum detectable vehicle | ~3 m length at PlanetScope resolution (presence only); ~1.5 m at SkySat/WorldView-3 (type classification feasible) |
| Cloud-cover constraint | Optical collection fails through thick cloud; no mitigation within optical-only workflow |
| Archive depth | Planet archive from 2016; Maxar archive from early 2000s for select sites; Pléiades from 2012 |
| Latency (routine monitoring) | 12–24 hours from acquisition to processed imagery via standard commercial pipelines |
| Delivery formats | GeoTIFF (imagery), GeoJSON / GeoPackage (vehicle count layers), PDF / structured report (analytic summary) |
| Off-nadir degradation threshold | Vehicle classification reliability degrades significantly beyond ~25 degrees off-nadir |
Analytics Satellize can run
| Vehicle presence count by type | Object detection via template matching and CNN-based classifier trained on confirmed vehicle imagery; shadow-length estimation for height-based disambiguation | GeoJSON layer with vehicle centroids, inferred class, and confidence score per acquisition |
| Throughput volume estimate | Queuing-theory model applied to vehicle count time series; dwell-time priors from published military logistics doctrine | Tabular report per facility: estimated vehicle-movements per day with stated uncertainty range |
| Anomaly alert on count surge | Rolling-baseline comparison; alert triggered when count exceeds baseline by configurable threshold (e.g. 2 standard deviations over 7-day window) | Automated alert with annotated image chip and count chart; delivered within 24 hours of triggering acquisition |
| Railcar activity time series | Linear feature detection to identify rail sidings; car-count change detection between successive acquisitions; empty vs. loaded discrimination by shadow profile | Time-series chart of railcar counts per siding with loaded/empty split; GeoPackage layer |
| Facility activity narrative | Analyst synthesis of automated count outputs, collection-gap flags, and contextual reporting; structured around named facilities and dated imagery | Periodic intelligence-format report (PDF) with image citations, confidence ratings, and gap disclosure |
| Multi-site comparative tempo index | Normalised vehicle-count index across a defined set of facilities; identifies which sites are showing elevated activity relative to their own historical baseline | Dashboard feed (GeoJSON or REST API) with per-site tempo index updated on each clear acquisition |
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.