Conveyor belt shadow and stockpile geometry flow-rate inference
Repeated stereo-derived surface models of mine stockpiles, combined with conveyor shadow geometry as an operational proxy, let analysts infer throughput flow rates between satellite passes without setting foot on site.
Sensors
- Maxar WorldView-3: 0.31 m panchromatic, 1.24 m multispectral. Collects same-pass stereo or tri-stereo pairs that support DSM generation at 0.5–1 m post spacing. Revisit at mid-latitudes roughly 1–4.5 days depending on tasking priority and off-nadir angle.
- Airbus Pléiades Neo: 0.30 m panchromatic, 1.2 m multispectral. Native stereo and tri-stereo acquisition in a single pass. Two-satellite constellation gives same-day revisit over most sites. DSM accuracy typically 0.5 m RMSE vertical in open terrain.
- DLR TanDEM-X: X-band bistatic SAR pair, 12 m posting in standard science mode, finer in High Resolution Spotlight (up to ~2 m). Cloud-independent. Bistatic coherence avoids temporal decorrelation; suited to change detection between tasked acquisitions separated by days to weeks.
- Planet Pelican: 0.30 m panchromatic, 0.50 m multispectral. Daily revisit globally. Resolution sufficient to detect conveyor shadow presence/absence and coarse stockpile shape change, though stereo DSM quality is inferior to WorldView-3 or Pléiades Neo for precise volumetrics.
What a stockpile's silhouette actually encodes
A conical ore stockpile has a predictable geometry: angle of repose for most crushed rock falls between 35° and 40°. Given that physical constraint, a plan-view footprint measured from sub-metre imagery translates directly into a volume estimate without needing a ground survey. When the same stockpile is imaged twice, the volume difference divided by the time between acquisitions gives a mean flow rate over that interval. The method is not new; it is the same principle used by port operators with terrestrial laser scanners, applied from orbit.
The honest limit is temporal averaging. A satellite pass captures one instant. If a conveyor ran at full capacity for six hours and sat idle for eighteen, the inferred daily rate will be correct on average but will not resolve the operational rhythm within the day. For that finer cadence, conveyor shadow geometry becomes the second data stream.
Conveyor shadows as a binary operational signal
An elevated conveyor belt gantry casts a shadow whose width and length are functions of solar geometry and belt height, both of which are known. What varies is whether material is piled on the belt: a loaded belt sits higher on its idler rollers and produces a subtly wider, occasionally asymmetric shadow compared with an empty one. At sub-0.5 m resolution this difference is detectable in favourable illumination conditions, though it is a probabilistic signal rather than a definitive measurement.
More reliably, the shadow reveals whether the belt structure itself is present and intact, and whether dust plumes at the transfer points are visible. Active transfer points generate fine particulate that scatters light and is visible in panchromatic imagery as a diffuse bright halo. Absent that halo, and with no change in stockpile geometry between passes, the inference is that the conveyor was idle. This is a proxy, not a direct measurement, and should be stated as such in any delivered product.
Planet Pelican's daily revisit is valuable here precisely because it trades volumetric precision for frequency. A sequence of daily 0.30 m images can establish which days showed active transfer-point dust signatures, giving the volume-change estimate a denominator that is closer to actual operating days than calendar days.
DSM generation and volume differencing: the numbers behind it
Stereo photogrammetry from WorldView-3 or Pléiades Neo produces a digital surface model using dense image matching. Over open, texturally rich terrain like crushed ore, vertical accuracy of 0.5–1.0 m RMSE is routinely achievable. For a stockpile 20 m tall and 80 m in diameter, a 1 m vertical uncertainty propagates to roughly 2–4% volume uncertainty, which is acceptable for production monitoring at the weekly to monthly scale.
TanDEM-X bistatic acquisitions offer a cloud-independent alternative. X-band penetrates only a few centimetres into dry ore, so the radar surface closely tracks the physical stockpile surface. The bistatic configuration means both antennas illuminate the target simultaneously, avoiding the temporal decorrelation that plagues repeat-pass InSAR over loose material. High Resolution Spotlight mode achieves approximately 2 m ground resolution, sufficient for stockpiles above roughly 50 m diameter. Smaller piles, or those with complex geometry from partial reclaim, require the optical stereo route.
Volume differencing requires a stable reference plane. The base of the stockpile pad, measured during a period when the pad is empty, serves as the datum. Errors in that datum propagate linearly into every subsequent volume estimate, so establishing it correctly at the outset is worth the effort.
Where the method breaks down
Cloud cover is the primary operational constraint for optical sensors. Major iron ore export terminals in Western Australia or Brazil's Pará state can see multi-week cloud-free windows, but wet-season operations in equatorial regions may force reliance on TanDEM-X for months at a time. Even then, TanDEM-X requires tasking coordination with DLR; it is not a free-running open constellation.
Stockpile blending complicates interpretation. Large terminals often run multiple ore grades simultaneously, reclaiming from one pile while building another. If the net volume change across the yard is small but individual piles have turned over completely, a yard-level volume difference will understate throughput. Tracking individual pile footprints separately, rather than treating the yard as a single volume, is necessary but requires careful manual or semi-automated pile segmentation.
Finally, the method infers throughput, not grade or commodity value. A pile of low-grade fines and a pile of high-grade lump ore look identical to a DSM. Spectral reflectance can offer partial discrimination, particularly for iron ore where Fe content correlates with near-infrared reflectance, but that is a separate analytic layer and carries its own uncertainties.
Putting it into practice: what a monitoring programme looks like
A workable programme for a large export terminal typically combines monthly Pléiades Neo stereo tasking for volumetrics with daily Planet Pelican passes for operational status flags. The stereo acquisition is planned around forecast cloud windows; the daily optical pass is opportunistic. TanDEM-X acquisitions are reserved for the wet season or as a cross-check against the optical DSMs twice a year.
Satellize structures analytics engagements around open constellations supplemented by commercial tasking on client licence, the same model used in its Tonga crop-estimation programme. For a mining client, that means the daily Planet layer runs on open archive while stereo tasking is procured per-acquisition against a defined site list. Deliverables are volume-change time series in GeoTIFF and CSV, with conveyor activity flags as a companion attribute table.
The output is most useful when paired with the client's own shipping manifest or rail-load data. Satellite-derived throughput and ground-truth logistics records will rarely agree to the tonne, but systematic divergence between them is exactly the kind of signal that justifies the exercise.
Typical figures
| Best optical resolution (panchromatic) | 0.30 m (WorldView-3, Pléiades Neo, Planet Pelican) |
| Stereo DSM vertical accuracy (open terrain) | 0.5–1.0 m RMSE (Pléiades Neo / WorldView-3 dense matching) |
| TanDEM-X ground resolution (High Resolution Spotlight) | ~2 m; standard science mode 12 m posting |
| Optical revisit (Planet Pelican) | Daily globally at 0.30 m; stereo pairs require tasking |
| Stereo tasking revisit (Pléiades Neo) | Same-day stereo possible; routine tasking 1–3 day lead time |
| Minimum stockpile diameter for reliable volume estimate | ~50 m for TanDEM-X HR Spotlight; ~20 m for optical stereo |
| Volume uncertainty (typical conical pile, 1 m DSM error) | 2–4% for piles >15 m tall; larger for low-profile reclaimed piles |
| Cloud dependency | Optical sensors fully cloud-blocked; TanDEM-X X-band cloud-independent |
| Archive depth (commercial optical) | WorldView-3 from 2014; Pléiades Neo from 2021 |
| Delivery formats | GeoTIFF DSM, CSV volume-change time series, GeoPackage pile footprints |
Analytics Satellize can run
| Stockpile volume time series | Dense stereo matching (SGM or similar) on sub-metre stereo pairs; DSM differencing against stable pad datum | Monthly GeoTIFF DSMs and CSV volume table per named pile, with uncertainty bounds |
| Inferred throughput flow rate | Volume delta divided by active operating days (from conveyor activity flags); angle-of-repose geometric constraint applied to footprint measurements | Tabular flow-rate estimates (tonnes per day equivalent, density-adjusted) per epoch |
| Conveyor operational status flags | Shadow geometry analysis and transfer-point dust halo detection in daily 0.30 m panchromatic imagery | Daily binary activity flag per conveyor segment, delivered as attribute table alongside imagery timestamps |
| Cloud-independent volume check (SAR) | TanDEM-X bistatic DSM differencing; X-band surface model compared against optical stereo baseline | Bi-annual SAR-derived volume cross-check report with residual analysis against optical estimates |
| Pile footprint segmentation layer | Supervised or threshold-based segmentation of DSM local maxima and slope breaks to delineate individual pile boundaries | GeoPackage polygon layer of individual pile extents, updated with each stereo acquisition |
| Logistics divergence alert | Comparison of satellite-derived throughput estimate against client-supplied shipping or rail manifest; statistical residual flagging | Monthly reconciliation report highlighting epochs where satellite and ground-truth estimates diverge beyond agreed tolerance |
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.