Cross-border rail wagon flow estimation at international rail crossings
High-resolution optical satellites count rail wagons queued or transiting at international border crossings, producing freight-flow proxies weeks before official statistics are published. The method is sharpest at gauge-break facilities where physical wagon transfers create dense, countable concentrations.
Sensors
- Maxar WorldView-3: Panchromatic resolution of 0.31 m native (delivered at 0.5 m under ITAR licensing) resolves individual wagon bodies and bogie assemblies clearly. Revisit at a given crossing is typically 1 to 4.5 days depending on latitude and tasking priority.
- Airbus Pléiades Neo: 0.3 m panchromatic, 1.2 m multispectral. Four-satellite constellation achieves same-day revisit with multiple look angles, useful for catching wagon movements within a single operational day. Stereo acquisition supports shadow-based length estimation.
- Planet SkySat: 0.5 m panchromatic, roughly 1 m multispectral. A fleet of around 20 satellites enables flexible tasking; video mode at up to 30 frames per second can capture wagon movement in near-real time, though coverage windows remain short.
- Sentinel-2 MSI: 10 m multispectral bands, 5-day revisit at mid-latitudes with the two-satellite constellation. Individual wagons are not resolved, but siding occupancy can be inferred from spectral and texture change at well-known facilities. Free and open; useful for historical baselining back to 2015.
Why a gauge-break yard is a natural counting point
At the boundary between the 1435 mm standard-gauge network used across most of Europe and the 1520 mm broad-gauge network inherited from the Soviet rail system, freight cannot simply roll through. Wagons must be transferred to different bogies, loaded onto flat-cars, or held in staging sidings while customs and phytosanitary checks proceed. The physical consequence is concentration: wagons that would otherwise be distributed across hundreds of kilometres of track are pooled at a handful of crossing facilities.
That pooling is precisely what makes satellite counting tractable. A crossing such as Brest-Terespol on the Poland-Belarus border, or Čierna nad Tisou on the Slovakia-Ukraine border, handles a large fraction of rail trade between the two networks. A count of wagons in the staging sidings on successive acquisition dates, differenced and divided by average wagon capacity for the commodity mix, yields a directional flow estimate. The method does not require the wagons to be moving at the moment of imaging; the queue itself is the signal.
What the detector actually sees, and what it can miss
At 0.3 to 0.5 m resolution, a standard European flat-car or covered hopper is roughly 14 to 20 m long and 3 m wide, spanning 30 to 70 pixels in the along-track dimension. That is enough for a convolutional object detector trained on labelled imagery to distinguish wagons from locomotives, maintenance vehicles, and fixed infrastructure. Covered hoppers, tank wagons, and flat-cars carrying intermodal units all have distinct roof signatures that allow coarse commodity-type classification.
The honest limits are several. Cloud cover is the most immediate: at Brest, average cloud frequency exceeds 60 percent in winter months, meaning a single acquisition may not be usable and multi-day compositing is sometimes necessary. Wagons parked under shed roofs or canopies are invisible. Wagons that transit the crossing between two acquisitions leave no trace in the count; the method captures stock, not flow, and flow must be inferred from stock changes. Finally, at facilities with very dense, multi-track sidings, wagon shadows at low sun angles can cause under-counting by obscuring adjacent rows.
Turning a count into a flow proxy
A single image gives an inventory. Two images, separated by a known interval, give a net change in wagon stock at the facility. If wagons are arriving from the east and departing westward, a rising eastern-siding count and falling western-siding count together indicate active westbound transfer. The arithmetic is straightforward; the difficulty is in assigning commodity types and load factors to the wagons counted.
Wagon-type classification from imagery is imperfect. Covered hoppers carrying grain look similar to those carrying potash. Tank wagons carrying oil look similar to those carrying chemicals. Where the commodity mix at a specific crossing is known from historical trade data or open customs sources, a weighted average net-weight-per-wagon figure can be applied. Where it is not, the output is a wagon-count index rather than a tonne estimate, which is still informative as a relative indicator of trade intensity across time.
Revisit cadence matters enormously for flow estimation. At a crossing processing several hundred wagons per day, a 24-hour gap between acquisitions captures one daily net change. A 5-day gap introduces ambiguity: wagons could have arrived and departed multiple times within the interval. For high-frequency crossings, combining SkySat or Pléiades Neo tasking with Sentinel-2 texture-change alerts to trigger commercial acquisitions is a practical architecture.
Archive depth and the geopolitical use case
Sentinel-2 data runs back to mid-2015, and Landsat 8 to 2013, giving nearly a decade of coarse-resolution change history at major crossings. While individual wagons are not resolved at 10 to 15 m, the spectral brightness and texture of densely occupied sidings differs measurably from empty ones, and a time-series of that signal can establish baseline seasonality and flag anomalous departures.
The geopolitical application is the most commercially sensitive. When official trade statistics between two countries are delayed, contested, or simply absent, satellite-derived wagon counts at border crossings offer an independent signal. Sanctions monitoring, treaty compliance verification, and commodity market intelligence all have potential interest in this data. The method is not a substitute for customs data but it is available faster, is not subject to the same reporting incentives, and covers crossings where official data may not be published at all.
Honest constraints a buyer should price in
Accuracy figures for wagon counting from VHR imagery are not yet standardised in the open literature in the way that, say, building footprint detection benchmarks are. Published studies on vehicle detection from satellite imagery (covering cars and trucks at parking facilities) report precision and recall figures in the 85 to 95 percent range under good conditions, but rail wagons at border crossings are a less-studied class with different occlusion patterns and a more constrained set of target shapes. Treat count outputs as having an uncertainty of roughly plus or minus 5 to 10 percent under good imaging conditions, widening significantly under partial cloud or low sun angle.
Latency from image acquisition to delivered count depends on tasking pipeline and processing setup. Same-day delivery is achievable with commercial tasking and automated detection; next-day is more typical in operational practice. Archive processing of historical imagery is faster than real-time tasking because cloud screening and scene selection can be done in batch.
Satellize runs detection pipelines on open and commercial constellations and can apply this method to specific crossing facilities on a scheduled or event-triggered basis.
Typical figures
| Spatial resolution (best available) | 0.3 m panchromatic (Pléiades Neo, WorldView-3); 0.5 m (SkySat) |
| Spatial resolution (open data) | 10 m multispectral (Sentinel-2); individual wagons not resolved |
| Revisit at a single crossing | Same-day feasible with Pléiades Neo four-satellite constellation; 1 to 4.5 days typical for WorldView-3 tasking; 5 days for Sentinel-2 at mid-latitudes |
| Minimum detectable target | Standard rail wagon (approx. 14 m × 3 m) clearly resolved at 0.5 m resolution; marginal at 1 m; not individually resolved at 10 m |
| Spectral bands used | Panchromatic (detection and counting); VNIR multispectral for wagon-type classification and shadow analysis |
| Cloud impact | Optical only; cloud cover above roughly 20 percent over the facility degrades or prevents counting. Multi-date compositing required at high-cloud-frequency sites |
| Archive depth | Sentinel-2: 2015 to present (free). Commercial VHR archives (WorldView, Pléiades): varies by site, typically 2008 to present for major crossings |
| Typical count uncertainty | Approximately ±5 to 10 percent under good conditions (clear sky, sun angle above 30°); wider under adverse geometry or partial occlusion |
| Delivery latency (operational) | Same-day to next-day for commercial tasking with automated detection pipeline; 24 to 48 hours typical |
| Output formats | GeoJSON feature layer (wagon centroids and type labels), CSV count time-series, georeferenced GeoTIFF with detections overlaid |
Analytics Satellize can run
| Wagon inventory count per siding zone | Convolutional object detection (sliding-window or anchor-based CNN) applied to pansharpened VHR imagery; wagon class trained on labelled examples from multiple crossing facilities | GeoJSON point layer with wagon centroids, type classification (covered, open, tank, flat), and confidence score; refreshed per acquisition |
| Net directional flow index (wagons per day) | Differencing of sequential inventory counts across eastbound and westbound staging zones; directional assignment based on facility zone mapping | CSV time-series with date, eastbound count, westbound count, net change, and data-quality flag (cloud fraction, sun angle) |
| Commodity-type composition estimate | Wagon-type classification from roof texture and shape signatures; weighted by known commodity-per-wagon factors from published rail freight statistics where available | Stacked bar chart report (PDF and data table) showing estimated commodity mix by wagon type per acquisition |
| Anomaly alert: unusual queue build-up or clearance | Statistical process control on rolling baseline wagon count; alert triggered when count exceeds two standard deviations from the seasonal mean | Email or API alert with image chip, count figure, historical context, and suggested follow-up tasking recommendation |
| Historical baseline and seasonality profile | Sentinel-2 siding-occupancy index (spectral brightness and NDVI-suppressed texture) computed monthly back to 2015; calibrated against VHR counts at overlapping dates | Multi-year time-series chart and GIS layer showing seasonal occupancy patterns and trend direction at the named crossing |
| Facility zone map (one-time setup) | Manual and semi-automated delineation of siding polygons, gauge-change shed footprints, and customs inspection areas from VHR base image | GeoPackage with labelled zone polygons used as the spatial framework for all subsequent count analytics |
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.