Iron ore stockpile grade estimation from spectral reflectance
Iron oxides absorb strongly near 900 nm and redden the visible slope. Hyperspectral satellites now read those signals from orbit, giving port and mine operators a non-contact grade proxy for surface stockpiles.
Sensors
- ASI PRISMA: Italian Space Agency hyperspectral imager, 30 m spatial resolution, 400–2500 nm range at approximately 10 nm spectral sampling, 30 km swath. Revisit roughly 29 days at nadir but programmable for tasking. The 850–950 nm region captures the ferric oxide absorption feature directly.
- NASA/JPL EMIT: Earth Surface Mineral Dust Source Investigation, mounted on the ISS. VSWIR coverage 380–2500 nm at ~7.4 nm spectral resolution, 60 m spatial resolution, 75 km swath. Designed for mineral mapping; its signal-to-noise ratio in the near-infrared is well suited to iron oxide characterisation. Coverage is ISS-track dependent, not fully programmable.
- Sentinel-2 MSI: 10–20 m spatial resolution, 5-day revisit at mid-latitudes. Band 8A (865 nm) and Band 4 (665 nm) straddle the ferric absorption feature coarsely. Useful as a change-detection layer and for stockpile boundary delineation, but 13 broad bands cannot resolve the absorption shape with the precision needed for grade regression. Treat as a coarse proxy only.
- AVIRIS-NG (airborne): NASA Airborne Visible/Infrared Imaging Spectrometer Next Generation. Spatial resolution adjustable from roughly 0.3 m to 4 m depending on flight altitude, 380–2510 nm at ~5 nm sampling. The gold standard for calibration and algorithm development, but campaign-based and not operationally scalable for routine stockpile monitoring.
What a 900 nm dip actually tells you
Iron in its oxidised forms, primarily haematite (Fe₂O₃) and goethite (FeOOH), produces two diagnostic spectral signatures. The first is a broad absorption centred near 900 nm driven by crystal-field transitions in the Fe³⁺ ion. The second is a steep rise in reflectance from blue to red wavelengths, which gives high-grade haematite ore its characteristic dark-red colour even to the naked eye. Together, the depth of the 900 nm feature and the slope of the visible reflectance curve correlate with iron content across a wide range of ore types.
Laboratory studies and published airborne campaigns have demonstrated that continuum-removed band depth at 900 nm, combined with a visible-to-NIR slope index, can estimate Fe₂O₃ content with root-mean-square errors in the range of 2–5 percentage points when models are trained on representative samples from the same deposit. That is not a substitute for assay, but it is useful for segregation decisions, blending optimisation and anomaly flagging at scale.
Where the physics breaks down
Moisture is the most serious confound. Water absorbs strongly at 970 nm and 1450 nm and suppresses overall reflectance, which compresses the apparent depth of the iron feature. A stockpile wetted by rain can appear lower-grade than it is. Corrections exist, using the 970 nm water absorption band as a moisture index, but they add uncertainty and require either concurrent moisture measurement or a dry-day acquisition strategy.
Particle size matters almost as much. Finely ground ore has a higher specific surface area, scatters more light, and appears brighter and less saturated than coarse lump ore of identical chemical composition. A model trained on fines will misestimate lump, and vice versa. Operators need to know the physical form of each stockpile before interpreting spectral results.
Atmospheric correction is non-trivial from orbit. PRISMA and EMIT both require per-scene correction for water vapour and aerosol path radiance before surface reflectance can be retrieved reliably. Dust above an active stockpile yard introduces additional uncertainty. Finally, the method reads only the surface layer, typically the top few millimetres. A stockpile that has been rained on, oxidised at the surface, or partially segregated during stacking may not represent the bulk grade.
From spectra to a grade map: the processing chain
The standard workflow begins with atmospheric correction to surface reflectance, followed by continuum removal across the 750–1000 nm window to isolate the absorption feature from background albedo variation. Band depth at the absorption minimum, integrated band area, and the asymmetry of the feature are then extracted as spectral indices. A regression model, typically partial least-squares regression or a support-vector machine trained on co-located assay data, maps those indices to estimated Fe content.
Spatial resolution matters here. PRISMA's 30 m pixels are adequate for large port stockpiles, which commonly span hundreds of metres, but will mix ore with concrete pad, shadow and equipment at stockpile edges. EMIT's 60 m pixels are coarser still. Sub-pixel unmixing, using spectrally pure endmembers for ore, shadow and substrate, can recover some spatial precision but introduces its own assumptions. For smaller stockpiles or finer segregation, airborne AVIRIS-NG data remain the practical benchmark.
Sentinel-2 as a monitoring backbone
PRISMA and EMIT acquisitions are infrequent and not always cloud-free. Sentinel-2, with its 5-day revisit and free archive back to 2015, fills the temporal gap as a change-detection layer. The ratio of Band 8A (865 nm) to Band 4 (665 nm) is a published proxy for iron oxide abundance, sometimes called the ferrous/ferric index, and it correlates broadly with haematite content. It cannot resolve the absorption shape, so it does not produce grade estimates with the precision of a hyperspectral model. What it does do is flag when a stockpile's spectral character has changed significantly between hyperspectral acquisitions, prompting a targeted re-task of PRISMA or a field check.
Stockpile boundary delineation is another Sentinel-2 strength. Consistent 10 m resolution in the visible bands allows automated mapping of stockpile footprints and height proxies from shadow geometry, providing the spatial mask needed before any spectral grade analysis is applied.
Operational constraints buyers should price in
Cloud cover is a hard blocker for all optical methods. Iron ore export terminals in tropical and sub-tropical regions, including Western Australia's Pilbara coast, Brazil's Pará state and Guinea's Simandou corridor, experience significant cloudy periods. A realistic planning assumption is that 30–50% of tasked acquisitions in wet-season months will be cloud-obscured at the moment of overpass. Combining multiple sensors and building a cloud-climatology model for the specific site is the only mitigation.
PRISMA is a research satellite operated by the Italian Space Agency. It is not a commercial constellation, so acquisition scheduling is subject to agency priorities and is not guaranteed on short notice. EMIT is similarly constrained by ISS orbital geometry and mission priorities. Buyers planning operational monitoring programmes should treat hyperspectral satellite data as high-value but irregular, and design workflows that degrade gracefully to Sentinel-2 proxies when hyperspectral data are unavailable.
Satellize structures analytics pipelines to handle exactly this kind of sensor hierarchy, switching between hyperspectral and multispectral inputs automatically and flagging confidence levels accordingly. The approach is similar in principle to the multi-source data fusion used in the Tonga crop-estimation programme, adapted here to the very different spectral physics of iron oxide mineralogy.
What the output is and is not
A well-calibrated hyperspectral stockpile grade map is a spatial index of relative iron content across a stockpile yard, updated at whatever cadence cloud and sensor availability permit. At its best, it can identify high-grade versus low-grade zones within a single large stockpile, support blending decisions before ship loading, and provide an independent check on declared grades for due-diligence or offtake-contract purposes.
It is not a replacement for physical sampling and laboratory assay. No remote-sensing method is. The honest use case is as a rapid, spatially continuous screening tool that directs physical sampling effort more efficiently, catches segregation anomalies that point-sampling would miss, and provides a documented spectral record of stockpile condition over time. That record has value for insurance, audit and dispute resolution, independent of any single grade estimate.
Typical figures
| Spatial resolution (hyperspectral) | 30 m (PRISMA), 60 m (EMIT) |
| Spatial resolution (multispectral proxy) | 10–20 m (Sentinel-2 MSI) |
| Spectral range | 400–2500 nm (PRISMA, EMIT, AVIRIS-NG); 443–2190 nm (Sentinel-2, 13 bands) |
| Spectral sampling (hyperspectral) | ~10 nm (PRISMA), ~7.4 nm (EMIT) |
| Revisit (hyperspectral) | ~29 days nadir (PRISMA, programmable); irregular ISS track (EMIT) |
| Revisit (Sentinel-2 proxy) | 5 days at mid-latitudes (2-satellite constellation) |
| Minimum stockpile size for reliable mapping | Approximately 5–10 ha for PRISMA; larger for EMIT; sub-hectare with AVIRIS-NG |
| Typical grade estimation error (published airborne studies) | 2–5 percentage points Fe₂O₃ RMSE with site-specific calibration |
| Archive depth | PRISMA from 2019; Sentinel-2 from 2015; EMIT from 2022 |
| Key confounds | Surface moisture, particle size variation, atmospheric aerosol, cloud cover, surface oxidation crust |
Analytics Satellize can run
| Per-stockpile Fe content index map | Continuum-removed band depth at ~900 nm plus visible-NIR slope regression, calibrated against client assay data; partial least-squares or support-vector regression | GeoTIFF grade-index raster per acquisition, with per-pixel uncertainty estimate |
| Stockpile boundary and footprint layer | Automated spectral segmentation on Sentinel-2 visible bands; shadow-geometry height proxy | GIS polygon layer (GeoJSON or Shapefile) updated on each cloud-free Sentinel-2 pass |
| Moisture-flagged acquisition report | 970 nm water absorption index applied to hyperspectral data to identify wet-surface pixels before grade interpretation | PDF acquisition quality report with moisture-affected pixel mask; recommendation on result reliability |
| Spectral change alert | Time-series analysis of Sentinel-2 Band 8A / Band 4 ratio across stockpile footprints; statistical threshold on deviation from rolling baseline | Automated alert (email or API push) when ratio shift exceeds threshold, prompting hyperspectral re-task |
| Historical spectral archive for due diligence | Retrospective processing of Sentinel-2 archive (2015-present) and available PRISMA scenes over named stockpile coordinates | Time-stamped spectral index time series as CSV and interactive chart; suitable for audit or offtake dispute documentation |
| Blending zone recommendation | Spatial clustering of grade-index map into high, medium and low zones; overlay with stockpile logistics geometry | Annotated stockpile map with suggested reclaim sequence zones; PowerPoint and GIS format |
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.