Thermal anomaly detection for spontaneous combustion in coal stockpiles and waste dumps
Spontaneous combustion in coal stockpiles and waste dumps is a safety, financial and air-quality hazard that conventional inspection misses. Thermal infrared satellites detect the surface expression of subsurface heating before visible smoke appears.
Sensors
- Landsat 8/9 TIRS: Two thermal infrared bands centred at 10.9 µm and 12.0 µm; native 100 m resolution resampled to 30 m in distributed products. Detects land surface temperature anomalies of roughly 0.5–1 K above background. 16-day revisit per satellite; combined Landsat 8 and 9 constellation gives 8-day repeat at mid-latitudes.
- ASTER TIR: Five thermal infrared bands (8.125–11.65 µm) at 90 m resolution, resampled to 15 m in some products. Better spectral discrimination than TIRS, useful for distinguishing combustion-related emissivity shifts from ambient warm ground. On-demand tasking only since 2008; archive extends to 1999.
- VIIRS I5 band (Suomi NPP / NOAA-20 / NOAA-21): Mid-wave infrared band centred at 11.45 µm, 375 m resolution, twice-daily overpass (day and night). Night-time passes maximise thermal contrast against cool background. Complements the VIIRS FIRMS active-fire detection chain and is particularly useful for persistent low-intensity smouldering that falls below MODIS detection thresholds.
- ECOSTRESS (ISS-mounted): Five TIR bands at approximately 70 m resolution, with irregular but sometimes sub-daily revisit depending on ISS orbital precession. Designed for evapotranspiration but equally sensitive to industrial heat anomalies. Useful for short-duration validation passes over known problem sites.
- Sentinel-3 SLSTR: Dual-view thermal infrared at 1 km nadir resolution; daily global coverage. Too coarse for individual stockpile mapping but useful for regional screening of large waste-dump complexes and for establishing multi-year baseline temperature climatologies.
Why coal ignites before anyone notices
Coal oxidises exothermically at ambient temperatures. When stockpiled material is finely crushed, poorly ventilated and left undisturbed, heat accumulates faster than it dissipates. The process accelerates above roughly 40°C and becomes self-sustaining well below the visible ignition point. A stockpile can be smouldering internally for weeks before surface temperatures rise enough to trigger a ground-level alarm.
Waste rock dumps present a related but distinct problem. Pyritic material in overburden oxidises via bacterial and chemical pathways, generating heat that can persist for decades. Some abandoned underground workings in major coal basins, including the Jharia field in India and parts of the Powder River Basin in the United States, have been burning continuously for years. The surface expression of these fires is often subtle: a patch of warm ground, a stress-cracked surface, a zone of anomalous vegetation die-off. Satellite thermal imaging is, in many cases, the only practical way to monitor them systematically.
What a thermal sensor actually measures, and what it misses
Thermal infrared sensors measure radiant energy emitted from the surface, which is then converted to a brightness temperature. The relationship between that reading and actual subsurface fire depth is indirect. A fire at 2 metres depth may produce a surface anomaly of 5–15 K above background; a fire at 10 metres may produce nothing detectable at all until it creates a vent or collapse feature. This is the fundamental limit of the method, and it is worth stating plainly: satellite thermal data cannot confirm the absence of deep-seated combustion.
Cloud cover blocks thermal infrared entirely. In monsoon-affected coal regions such as eastern India or parts of Indonesia, usable clear-sky acquisitions can be sparse for months at a time. The practical workaround is to build a dense multi-year archive so that each cloud-free pass is interpreted against a well-established seasonal baseline rather than a single prior image. Landsat's 40-plus-year archive makes this feasible for most major coal-producing regions. Night-time VIIRS passes reduce the solar-reflection contamination that complicates daytime thermal interpretation, and they often show cleaner anomaly contrast because ambient surface temperatures drop while subsurface heat sources continue to radiate.
Detection thresholds and the physics of finding a hotspot
Landsat TIRS has a noise-equivalent temperature difference (NEdT) of approximately 0.4 K under typical conditions. That sounds precise, but the practical detection floor for a discrete hotspot is higher because the anomaly must occupy a significant fraction of the 100 m pixel to register above background variability. A hotspot covering less than roughly 10–20% of a pixel is likely to be diluted below detection. This means small, early-stage combustion zones in a large stockpile may go undetected until they grow.
ASTER's five TIR bands allow temperature-emissivity separation, which is analytically useful. Coal and combustion residues have distinct emissivity signatures in the 8–12 µm range, and multi-band analysis can help distinguish genuine thermal anomalies from warm bare rock or metal roofing that happens to be sun-heated. The VIIRS FIRMS system, originally designed for vegetation fire detection, applies a contextual algorithm that flags pixels significantly warmer than their neighbours. Applied to industrial sites, the same logic works well for persistent smouldering, though the 375 m resolution means a single flagged pixel could represent a very large area of uncertainty.
Building a monitoring programme that holds up operationally
A credible thermal monitoring programme for a coal facility needs three things: a seasonal baseline, a change-detection rule, and a triage protocol for ground verification. The baseline is built from cloud-free Landsat and VIIRS acquisitions over at least two full years, stratified by season and time of day. Without it, a warm summer afternoon reading is indistinguishable from a genuine combustion anomaly.
Change detection is typically implemented as a threshold on the difference between observed brightness temperature and the seasonal climatological mean for each pixel, expressed in standard deviations. A common operational trigger is 3 sigma above the local background, though the appropriate threshold depends on site-specific variability and the cost of false alarms versus missed detections. Alerts above threshold are ranked by anomaly magnitude and spatial persistence across consecutive passes before any ground team is dispatched. A hotspot that appears in a single night-time VIIRS pass but not in the subsequent Landsat overpass is more likely a transient industrial process than a developing fire.
Regulatory and insurance context
Several jurisdictions now require operators of large coal stockpiles and waste dumps to maintain documented evidence of thermal monitoring as part of their environmental and safety licences. Satellite-derived thermal records, when properly archived with acquisition metadata and processing provenance, satisfy this requirement in a way that periodic manual inspections do not. They provide a continuous, independently verifiable record that is difficult to dispute.
For insurers writing property or business-interruption cover on coal handling facilities, a multi-year thermal anomaly record changes the underwriting conversation. A site with three years of clean thermal history and a documented alert-response protocol is a materially different risk from one with no systematic monitoring at all. Satellize incorporates this kind of longitudinal archiving into its analytics delivery, drawing on the same open-constellation data pipelines used in its Tonga crop-estimation programme. The deliverable is a signed, versioned data product rather than a screenshot.
Honest limits and the role of complementary data
To summarise the constraints plainly: thermal satellites detect surface expression only, cloud blocks acquisition, spatial resolution dilutes small hotspots, and deep fires with no surface vent are invisible. These are not reasons to avoid the method. They are reasons to pair it with complementary inputs.
Gas sensors, whether ground-mounted or airborne, detect CO and CO2 plumes that precede visible surface heating. InSAR deformation data (covered separately in the ground-subsidence page) can reveal the surface collapse that often accompanies a maturing underground fire. Optical imagery at 30–50 cm resolution, available commercially, can identify the surface cracking and discolouration that thermal sensors flag but cannot characterise. Used together, these layers give an operator a defensible picture of combustion risk. Used alone, thermal satellite data is a screening tool, not a definitive diagnosis.
Typical figures
| Thermal spatial resolution | ASTER TIR: 90 m native; Landsat 8/9 TIRS: 100 m native (30 m resampled); VIIRS I5: 375 m; ECOSTRESS: ~70 m; Sentinel-3 SLSTR: 1 km nadir |
| Revisit frequency | Landsat 8+9 combined: ~8 days at mid-latitudes; VIIRS: twice daily (day and night); ECOSTRESS: irregular, potentially sub-daily; Sentinel-3: daily global |
| Spectral bands used | Thermal infrared 8–12 µm (TIR); VIIRS I5 centred at 11.45 µm; ASTER bands 10–14 covering 8.125–11.65 µm |
| Temperature sensitivity (NEdT) | Landsat TIRS: ~0.4 K; ASTER TIR: ~0.3 K; practical anomaly detection floor varies with hotspot size relative to pixel area |
| Minimum detectable hotspot | Approximately 10–20% pixel fill required for reliable detection; sub-pixel hotspots detectable only if temperature contrast is very large (>50 K above background) |
| Cloud penetration | None. TIR is blocked by cloud. SAR (not thermal) is required for cloud-immune surface monitoring. |
| Archive depth | Landsat: 1972–present (TIR from Landsat 4, 1982); ASTER: 1999–present; VIIRS: 2012–present |
| Latency (open data) | Landsat: typically 6–12 hours post-acquisition via USGS EarthExplorer; VIIRS FIRMS: near-real-time, typically 3 hours |
| Delivery formats | GeoTIFF brightness-temperature grids, anomaly shapefiles, time-series CSV, alert feeds (GeoJSON or email), PDF monitoring reports |
| Coverage | Global; no tasking required for Landsat or VIIRS (push acquisition); ASTER and ECOSTRESS require scene selection from existing archive or scheduled tasking |
Analytics Satellize can run
| Thermal anomaly baseline map | Seasonal climatology built from multi-year Landsat and VIIRS clear-sky acquisitions; per-pixel mean and standard deviation by month and overpass time | GeoTIFF raster stack and summary PDF; updated annually or on request |
| Change-detection alert layer | Z-score threshold (typically 3 sigma above seasonal mean) applied to each new clear-sky acquisition; persistent anomalies flagged across consecutive passes to reduce false positives | GeoJSON alert feed with anomaly magnitude, pixel coordinates and confidence tier; delivered within 24 hours of cloud-free acquisition |
| Hotspot growth or decay time series | Per-anomaly tracking across all available thermal acquisitions; area and peak brightness temperature extracted at each timestep | CSV time series and interactive chart; updated with each new acquisition |
| Night-time vs daytime anomaly comparison | VIIRS night-time I5 passes compared against daytime Landsat TIRS to separate solar-heated surfaces from genuine subsurface heat sources | Classified raster distinguishing probable combustion anomalies from diurnal solar artefacts; included in monthly monitoring report |
| Multi-sensor fusion confidence score | Logical combination of VIIRS flag, Landsat anomaly magnitude and (where available) ECOSTRESS or ASTER confirmation; Bayesian-style confidence tier assigned to each flagged zone | Attributed shapefile with confidence tier (low / medium / high) and sensor provenance for each flagged polygon |
| Regulatory compliance archive | Versioned, metadata-rich storage of all processed thermal products with acquisition timestamps, processing parameters and analyst sign-off | Signed PDF summary reports and archived GeoTIFF products suitable for submission to environmental regulators or insurers |
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.