Rail wagon and intermodal yard occupancy from satellite imagery
Track geometry forces rolling stock into predictable linear positions, making satellite-based counting far more tractable than open-area vehicle detection. High-resolution optical and SAR data can count, classify and monitor loading state across classification yards and intermodal terminals globally.
Sensors
- Sentinel-1 SAR (C-band, ESA): Interferometric Wide Swath mode delivers 10 m ground range resolution with a 6-day repeat at mid-latitudes (12-day single-pass). Cloud-independent, day-night capable. VV and VH polarisation channels allow backscatter contrast between empty open wagons (low return) and loaded wagons carrying bulk ore or scrap (high return). Free to access.
- Planet SuperDove (PlanetScope): 3 m native resolution, 8 spectral bands, near-daily revisit at most latitudes. Sufficient to resolve individual wagon lengths (standard freight wagon: 12–25 m) and distinguish wagon type by roof geometry. Cloud cover remains a hard limit; high-latitude winter performance degrades significantly.
- Capella Space SAR (X-band): Spotlight mode achieves 0.5 m resolution, enabling individual axle-group discrimination and confident classification between flat-cars, tank wagons and covered hoppers even in full darkness or cloud. Tasked commercially; latency depends on contract tier but same-day delivery is published as a product.
- COSMO-SkyMed (X-band, ASI): Spotlight mode at approximately 1 m resolution. Operated by the Italian Space Agency with commercial access via e-GEOS. Useful for high-confidence wagon-type classification where Capella capacity is unavailable. Constellation of four satellites gives revisit of roughly 12 hours at mid-latitudes under ideal tasking.
Why track geometry does half the work for you
A freight wagon cannot be anywhere. It sits on one of a finite set of tracks, oriented along a known azimuth, at a position constrained by the wagon immediately in front of it. This is fundamentally different from counting trucks in a car park, where targets scatter across two dimensions. In a classification yard, the search space collapses to a set of one-dimensional strings. Once track centrelines are mapped from a single high-resolution baseline image, subsequent analysis needs only to scan those lines for occupied versus unoccupied segments.
Track centreline extraction is a solved problem at 0.5–3 m resolution. At 10 m (Sentinel-1), individual tracks in a dense yard merge, but the yard as a whole produces a distinctive linear backscatter signature that separates clearly from surrounding land cover. The practical upshot: even moderate-resolution SAR can give reliable yard-level occupancy indices, while high-resolution optical or spotlight SAR is needed for wagon-level counting and type classification.
What a wagon's roof shape gives away
Wagon classification from above relies almost entirely on roof geometry and spectral signature. Flat-cars carrying intermodal containers show the rectangular top surface of the container, typically in a high-reflectance colour. Empty flat-cars show the dark steel deck. Tank wagons present a curved, specularly reflective surface that in SAR produces a characteristic double-bounce return from the cylinder-rail interface. Covered hoppers have a distinctive ridged roof profile visible at sub-metre resolution.
At Planet SuperDove's 3 m resolution, individual wagon roofs are resolvable but type classification is ambiguous for shorter wagon classes. A convolutional approach trained on higher-resolution imagery and applied to 3 m data can push classification accuracy up, but honest expectation-setting matters: confident type discrimination at 3 m requires the wagon to be at least 15 m long and the image to be acquired at low off-nadir angle. Capella's 0.5 m spotlight mode removes most of that ambiguity.
SAR polarimetry and the loaded-versus-empty question
Open wagons, gondolas and hoppers used for bulk commodities present a measurable backscatter difference between loaded and empty states in C-band and X-band SAR. An empty steel gondola is essentially a corner reflector: the vertical walls and flat floor produce strong double-bounce returns in VV polarisation. Fill that gondola with iron ore or coal and the rough, absorbing surface suppresses the double-bounce and increases diffuse scattering, shifting the VV/VH ratio. Published research using Sentinel-1 data has demonstrated this effect is detectable at the individual wagon scale when 1 m or better resolution SAR is used, and at the consist scale with 10 m data by averaging across a known wagon string.
The method has limits. Covered hoppers and tank wagons show no useful loading-state signal from above because the commodity is enclosed. Flat-cars carrying containers are trivially classified as loaded by container presence. The loading-state signal is therefore most useful for open bulk wagons, which happen to be economically important: coal, ore and grain movements are precisely the flows that commodity traders and infrastructure planners want to monitor.
High-latitude yards and the seasonal revisit problem
Optical sensors are useless under cloud, and high-latitude yards spend a substantial fraction of winter under persistent overcast. A yard at 60°N can expect cloud cover exceeding 70% of days between November and February. Daily optical revisit becomes, in practice, a useful image every week or two during winter months. For yards in Siberia, northern Canada or Scandinavia, this is not an edge case; it is the operational baseline.
SAR is the answer, and it is not a compromise. Sentinel-1's 6-day repeat at 60°N, combined with Capella or COSMO-SkyMed tasking on priority dates, can sustain a cadence of two to four observations per week regardless of weather. The trade-off is that SAR-only analysis cannot reliably classify wagon type without sub-metre resolution, so the analytic product shifts from type-level counts to occupancy indices and loading-state estimates. Clients operating in high-latitude regions should plan their data architecture around SAR as the primary layer and optical as an opportunistic supplement, not the reverse.
Turning counts into commercial intelligence
Raw wagon counts become useful when they are time-stamped, georeferenced to individual tracks or yard zones, and set against a baseline. A yard normally holding 400 wagons that shows 600 over three consecutive passes is accumulating dwell. That accumulation is a congestion signal, a demand signal, or evidence of a service disruption, depending on context. Dwell-time distributions derived from multi-pass imagery have been used in published academic work to infer network bottlenecks without access to any operator data.
Intermodal terminals add a layer of complexity because the relevant unit is the container on the flat-car, not the wagon itself. Container presence, stack height (estimable from shadow length at known sun angle) and terminal zone occupancy are all derivable from sub-metre optical or SAR data. Satellize structures analytics of this type as time-series feeds rather than one-off snapshots, which is how the Tonga crop-estimation programme is also delivered: repeated observations against a stable baseline, with change flagged automatically.
The honest ceiling: satellite analytics cannot read wagon identification numbers at any commercially available resolution, cannot determine commodity type inside a closed wagon, and cannot substitute for waybill data when origin-destination flows are the question. What they can do is provide an independent, non-cooperative observation of physical state across any yard on Earth, at a cadence and cost that ground surveys cannot match.
Typical figures
| Spatial resolution (optical) | 3 m (Planet SuperDove); sub-metre available via commercial tasking |
| Spatial resolution (SAR) | 10 m Sentinel-1 IW; 0.5 m Capella spotlight; ~1 m COSMO-SkyMed spotlight |
| Revisit cadence | 6-day Sentinel-1 (single pass); near-daily Planet at mid-latitudes; same-day tasking via Capella on contract |
| SAR frequency | C-band 5.4 GHz (Sentinel-1); X-band 9.6 GHz (Capella, COSMO-SkyMed) |
| Minimum detectable wagon length | ~12 m at 3 m optical resolution; ~6 m at 0.5 m SAR resolution |
| Cloud penetration | Full (SAR); zero (optical) |
| Archive depth | Sentinel-1: from 2014; Planet: from 2016 at most sites; Capella: from 2020 |
| Latency (operational product) | 3–6 hours post-acquisition for SAR (Sentinel-1 NRT); same-day for tasked commercial SAR |
| Typical coverage per pass | Sentinel-1 IW swath: 250 km wide; Planet scene: ~25 km × 23 km tile |
| Delivery format | GeoTIFF count layers, GeoJSON wagon polygons, time-series CSV, API feed |
Analytics Satellize can run
| Yard occupancy index | Track-constrained object detection on SAR backscatter intensity, aggregated to yard zones defined from baseline optical imagery | Daily or per-pass index (0–100%) per named yard zone, delivered as time-series CSV or dashboard feed |
| Wagon count by type | Roof-geometry classification using convolutional neural network trained on sub-metre optical and spotlight SAR; applied to Planet or Capella tasked imagery | GeoJSON polygon layer with wagon centroid, estimated type (flat-car, tank, hopper, gondola) and confidence score |
| Loading-state estimate for open wagons | VV/VH backscatter ratio analysis on Sentinel-1 or Capella SAR; threshold derived from published double-bounce suppression effect for bulk commodity fill | Per-consist loaded/empty flag appended to wagon count layer; reported with confidence interval |
| Dwell-time distribution | Multi-pass change detection: wagon string persistence across sequential observations used to estimate minimum dwell; method follows published network-bottleneck inference approaches | Weekly histogram of estimated dwell by yard zone; anomaly alert when median dwell exceeds configurable threshold |
| Intermodal terminal container occupancy | Container detection from shadow-length and roof-colour signature in high-resolution optical; stack-height estimation from sun-elevation geometry | GeoTIFF heatmap of terminal occupancy density; zone-level fill-rate report |
| High-latitude winter monitoring package | SAR-primary fusion workflow combining Sentinel-1 baseline with Capella tasking on priority dates; optical imagery ingested opportunistically when cloud-free | Continuous occupancy index feed with data-source flag per observation; no data gaps flagged explicitly rather than interpolated |
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.