Processing plant throughput inference from thermal and stockpile change
Combining Landsat-9 thermal radiance over crushers and mills with stereo-derived stockpile volume changes produces a credible, independent estimate of ore processing rates, without setting foot on site.
Sensors
- Landsat-9 TIRS Band 10: Thermal infrared at 10.6–11.19 µm, 100 m native resolution (resampled to 30 m in products), 16-day revisit at the equator. Measures at-surface brightness temperature to roughly ±0.3 K after atmospheric correction, sufficient to resolve the elevated skin temperature of a loaded SAG mill roof or crusher building against ambient ground.
- Planet SuperDove (PlanetScope): Eight-band multispectral at 3–4 m GSD, near-daily revisit globally. Used here for stockpile footprint delineation and surface change detection rather than thermal sensing. Shadow-geometry analysis of stockpile edges supports volume estimation when stereo pairs are unavailable.
- Maxar WorldView Legion: 30 cm panchromatic, sub-daily revisit in high-latitude tasking modes. Stereo pairs from the same orbital pass yield DEMs with 0.3–0.5 m vertical accuracy under good contrast conditions, enabling precise stockpile volume differencing between acquisition epochs.
- Airbus Pléiades Neo: 30 cm panchromatic stereo and tri-stereo capability, 50 cm multispectral. Tri-stereo acquisitions reduce occlusion artefacts on the steep flanks of ore stockpiles, improving volume estimates. Tasking latency is typically 1–3 days from order.
What a loaded mill radiates that an idle one does not
A semi-autogenous grinding mill running at full throughput dissipates several megawatts of heat through friction, motor losses and ore-on-steel contact. That energy has to go somewhere. A significant fraction conducts through the mill shell and radiates from the roof and walls of the enclosing structure, raising surface brightness temperature by 3–8 K above ambient, depending on insulation, ambient air temperature and load factor. Landsat-9 TIRS Band 10 resolves temperature differences of roughly 0.3 K after atmospheric correction, so a loaded mill is not subtle.
The same logic applies to primary crushers, secondary crushers and conveyor drive stations, each of which has a characteristic thermal signature under load. Published work on cement kilns and electric arc furnaces uses precisely this approach: correlating TIRS-derived brightness temperature with production logs to build a calibration curve. The mining application is analogous, though the absolute temperatures are lower and the spatial footprint smaller, which is why 100 m TIRS pixels require careful sub-pixel unmixing when multiple heat sources sit within a single pixel.
Stockpile geometry as an independent check
Thermal signatures tell you the plant is running. They do not, on their own, tell you at what rate ore is moving through it. Stockpile volume change provides the second channel. If the run-of-mine pad is shrinking and the product stockpile is growing, the plant is processing. The rate of volume change, divided by bulk density (typically 1.6–2.2 t/m³ for crushed hard rock, depending on ore type), converts directly to mass throughput.
Stereo DEM differencing is the standard method. A WorldView Legion or Pléiades Neo stereo pair acquired at time T1 is subtracted from a pair acquired at T2. Stockpile volumes are extracted by differencing the DEM against a baseline ground surface. Vertical accuracy of 0.3–0.5 m from commercial stereo at 30 cm GSD translates to volume uncertainty of roughly 5–10% for a stockpile of moderate size, say 50,000 to 200,000 tonnes. Smaller piles carry proportionally larger relative error.
Planet SuperDove imagery fills the gaps between stereo acquisitions. At 3–4 m resolution it cannot resolve fine topography, but shadow-length photogrammetry and footprint-area tracking can detect significant pile growth or drawdown between stereo epochs, flagging when a fresh stereo acquisition is warranted.
Honest limits: what the method cannot resolve
Cloud cover is the persistent adversary. Landsat-9 TIRS cannot penetrate cloud, and stereo DEM generation fails over cloud shadow. In tropical mining districts, cloud-free Landsat acquisitions may arrive only four to six times per year. Combining Landsat with ASTER, Sentinel-3 SLSTR (1 km thermal, daily) and commercial thermal tasking where available improves temporal coverage but does not eliminate the problem.
The thermal signal is ambiguous between throughput and ore type. A mill grinding harder, more abrasive ore at the same feed rate runs hotter than one processing softer material. Without ore-type context, the thermal proxy can overestimate throughput when feed hardness changes. Similarly, ambient temperature swings of 15–20 K between seasons shift the baseline and must be removed before the anomaly is meaningful.
Stockpile volume estimates assume stable bulk density. In practice, moisture content and degree of compaction vary, introducing errors that can reach 15% in wet conditions. These are not fatal to the method, but they set a realistic floor on precision: expect throughput estimates accurate to roughly ±15–25% under favourable conditions, not ±5%.
Calibration: turning radiance into tonnes
A raw thermal anomaly is a dimensionless signal until it is tied to a reference. Calibration uses periods when reported production figures are available, either from public filings (quarterly production reports are mandatory for listed miners in most jurisdictions) or from operational data shared under a client agreement. Regressing reported throughput against the combined thermal-plus-stockpile signal over those periods produces a site-specific transfer function. Once established, the function can be applied forward to estimate throughput in periods where no reported figure exists.
Transfer functions are not universal. A copper concentrator and a gold heap-leach facility have different thermal footprints and different stockpile geometries. Each site requires its own calibration, ideally spanning at least 12 months to capture seasonal variation in ambient temperature and ore hardness. Sites with publicly available monthly production data, common among ASX and TSX-listed juniors, are the easiest to calibrate quickly.
Where the combined signal earns its keep
The most direct application is independent production verification for commodity traders, lenders and royalty holders who receive throughput reports from operators but have no means to audit them. A satellite-derived estimate, even at ±20% accuracy, is informative when it diverges significantly from a reported figure.
A second application is competitive intelligence for mining companies benchmarking a rival's ramp-up after commissioning. Public production guidance is often vague in the first year of operation. The thermal and stockpile signal resolves whether a new mill is running at 60% or 90% of nameplate capacity months before a quarterly report confirms it.
Satellize applies this dual-channel method as part of its broader mining-site analytics capability, drawing on the same open-constellation and commercial-tasking infrastructure it uses for programmes such as the Kingdom of Tonga crop-estimation work. The method is most useful when paired with equipment-count analysis from optical imagery, which provides a separate activity proxy, though that sits on a sibling page.
Revisit, latency and what a practical monitoring cadence looks like
A workable operational cadence combines Landsat-9 thermal at 16-day repeat with Planet SuperDove daily optical for stockpile footprint tracking, and commercial stereo tasked monthly or on trigger. Landsat-9 archive extends to August 2021, Landsat-8 to February 2013, giving nearly 12 years of thermal history at consistent radiometric calibration. That archive depth allows retrospective analysis of a plant's operating history, useful for due diligence on an acquisition target.
Alert latency from a Landsat pass to a processed thermal anomaly report is typically 24–48 hours using standard USGS Collection 2 Level-2 products, which include surface temperature estimates derived from the single-channel algorithm. Stereo DEM delivery from commercial providers runs 3–7 days from tasking, depending on cloud and provider queue. The combined product, a throughput estimate with uncertainty bounds, is realistic at roughly weekly cadence in clear-sky conditions and monthly in cloud-prone regions.
Typical figures
| Thermal spatial resolution | 100 m native (Landsat-9 TIRS Band 10), resampled to 30 m in USGS Level-2 products |
| Thermal temperature sensitivity | ~0.3 K noise-equivalent temperature difference after atmospheric correction (Landsat-9 TIRS) |
| Optical resolution for stockpile delineation | 3–4 m (Planet SuperDove); 30 cm (WorldView Legion / Pléiades Neo stereo) |
| Stereo DEM vertical accuracy | 0.3–0.5 m (WorldView Legion / Pléiades Neo at 30 cm GSD, good contrast) |
| Landsat-9 revisit | 16 days at equator; combined with Landsat-8, effective 8-day repeat |
| Planet SuperDove revisit | Near-daily globally (constellation of ~200 satellites) |
| Throughput estimate accuracy | ±15–25% under favourable conditions after site-specific calibration |
| Minimum detectable thermal anomaly | ~3 K above ambient at 100 m pixel scale; sub-pixel unmixing required for sources <1 ha |
| Landsat thermal archive depth | Landsat-8 from February 2013; Landsat-9 from August 2021 (consistent TIRS calibration) |
| Alert latency (thermal) | 24–48 hours from satellite pass using USGS Collection 2 Level-2 surface temperature products |
Analytics Satellize can run
| Monthly throughput index | Regression of TIRS brightness temperature anomaly and stereo-derived stockpile volume change against reported production data; site-specific transfer function applied forward | Monthly time-series report with point estimate and ±1σ uncertainty bounds, delivered as PDF and CSV |
| Thermal anomaly alert | Automated z-score detection on TIRS Band 10 pixel values within plant boundary mask; triggered when anomaly exceeds 2σ from 90-day rolling baseline | Email or API alert within 48 hours of Landsat overpass, with annotated scene thumbnail |
| Stockpile volume change layer | Stereo DEM differencing (WorldView Legion or Pléiades Neo); baseline ground surface from earliest available stereo epoch; volume extracted per named pile polygon | GeoTIFF DEM difference raster and polygon-level volume table, updated per stereo acquisition |
| Ramp-up trajectory assessment | Time-series of thermal anomaly magnitude and stockpile drawdown rate from commissioning date; compared against nameplate capacity benchmark | Single-page ramp-up assessment report with annotated chart, suitable for investment committee use |
| Historical production reconstruction | Retrospective TIRS analysis over Landsat-8/9 archive (2013–present); calibrated against public quarterly filings where available | Annual archive of monthly throughput estimates as GIS-ready time-series layer and tabular data |
| Production divergence flag | Comparison of satellite-derived throughput estimate against operator-reported figures from public filings; divergence flagged when satellite estimate falls outside ±30% of reported value | Quarterly divergence report with confidence classification (high / medium / low) and supporting evidence imagery |
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.