Steel mill production activity monitoring by thermal anomaly detection

Active blast furnaces and electric-arc furnaces produce intense, spatially compact thermal anomalies detectable in shortwave and mid-wave infrared bands. Satellite sensors from Landsat to VIIRS can track operational status and relative throughput, giving commodity analysts and insurers an independent production signal.

Sensors

Landsat 8/9 TIRS: Two thermal infrared bands centred at 10.9 µm and 12.0 µm, 100 m native resolution (resampled to 30 m). Sixteen-day repeat at any given point, though combined Landsat 8 and 9 operations halve that to roughly eight days. Radiometric saturation occurs above approximately 750 K, which can clip the hottest furnace pixels; sub-pixel radiance mixing still yields a usable anomaly signal.
ASTER TIR: Five thermal bands from 8.1 to 11.7 µm at 90 m resolution. On-demand tasking historically available; instrument has operated in a reduced-acquisition mode since 2011 but archived scenes over major industrial sites are extensive. Better spectral discrimination than TIRS for separating coke-oven and furnace signatures.
VIIRS I4 / I5 bands: I4 at 3.74 µm (mid-wave infrared) and I5 at 11.45 µm (thermal infrared), both at 375 m resolution. Suomi NPP and NOAA-20 together provide at least one daytime and one night-time overpass daily. Coarse resolution blurs individual furnaces at most sites but the nightly cadence makes VIIRS the best tool for detecting abrupt shutdowns or restarts within a 24-hour window.
Planet SkySat: Shortwave-infrared-adjacent bands in the near-infrared at sub-metre to 0.5 m panchromatic resolution. Not a dedicated thermal sensor, but published research has used elevated NIR radiance from very high-temperature sources as a proxy anomaly indicator when true SWIR is unavailable. Tasking on demand; useful for corroborating furnace location and site layout at high spatial detail.
MODIS Terra / Aqua (legacy context): 500 m and 1 km bands including Band 21/22 at 3.96 µm, designed for fire detection but sensitive to industrial thermal anomalies. Daily global coverage. Largely superseded by VIIRS for new work but the archive back to 2000 is valuable for long baseline analysis of mill operating history.

What a furnace tap looks like from 700 km up

A blast furnace tap releases molten iron at roughly 1,480 to 1,520 °C. That temperature produces radiance in the shortwave infrared (SWIR, 1.5 to 2.5 µm) and mid-wave infrared (MWIR, 3 to 5 µm) that is orders of magnitude above the background signal from warm steel structures or sun-heated roofing. Even when the furnace itself is enclosed, the tap hole, the torpedo ladle car beneath it, and the slag runner all radiate intensely for the duration of the cast, typically 90 to 150 minutes for a large blast furnace.

At Landsat's 100 m thermal resolution, a single furnace tap occupies a fraction of one pixel. But radiance mixing still inflates that pixel's brightness temperature measurably above background, and the anomaly is reproducible across passes. Electric-arc furnaces present a different signature: shorter, more frequent heats (roughly 40 to 60 minutes per heat for a modern EAF), higher peak temperatures at the arc itself, and a characteristic pattern of thermal pulses rather than the slower, steadier signal of a continuously operating blast furnace.

Turning a temperature spike into a production index

The core method is sub-pixel thermal anomaly detection. For each satellite overpass, analysts compute brightness temperature or at-sensor radiance for every pixel covering the mill footprint, then compare it against a rolling baseline derived from cloud-free historical passes. A pixel that exceeds the baseline by a statistically significant margin, and that sits within the known furnace location, registers as an active anomaly. The spatial compactness of the anomaly (one or two pixels, not a diffuse warm zone) distinguishes a furnace from, say, a sun-heated ore stockpile.

Aggregating anomaly detections over time produces a binary on/off operational record per furnace. With sufficient overpass frequency, analysts can estimate relative utilisation: a mill that shows a thermal anomaly in 80 percent of cloud-free Landsat passes over a quarter is almost certainly running at higher capacity than one showing anomalies in 30 percent. This is not a direct tonne-per-day figure. It is a relative index, and it is most useful when tracked against itself over time or compared across competing mills in the same region.

Calibration against publicly reported production data from major producers allows analysts to anchor the index to physical units, at least approximately. The relationship is noisy. Furnace size, product mix, and tap scheduling all introduce variance that satellite data alone cannot resolve.

Where the method breaks down

Cloud cover is the most obvious limit. Thermal infrared does not penetrate thick cloud. In regions with persistent overcast, such as parts of coastal China or the Great Lakes steel belt in winter, cloud-free Landsat passes may be weeks apart. VIIRS's nightly cadence helps, but at 375 m a single VIIRS pixel covers an area that may contain multiple furnaces, a coke battery, and a power plant simultaneously. Attribution to a specific production unit becomes speculative.

Sensor saturation is a genuine problem at the hottest sources. Landsat TIRS Band 10 saturates at brightness temperatures above roughly 750 K. A tap hole during active casting can exceed this locally, clipping the signal and preventing quantitative radiance retrieval for that pixel. The anomaly is still detectable as a saturated flag, but the magnitude information is lost. ASTER's narrower dynamic range makes it similarly susceptible.

Coke ovens, which sit adjacent to blast furnaces at integrated steelworks, produce their own persistent thermal signatures. Separating coke-oven emissions from blast-furnace taps requires either higher spatial resolution, spectral discrimination across multiple thermal bands, or careful use of site layout knowledge to mask known coke-battery locations. At coarse VIIRS resolution, this separation is often not possible. Product grade, steel chemistry, and downstream processing steps are entirely invisible to thermal sensors.

The financial intelligence case

Steel production data from national statistical agencies is typically released with a one- to three-month lag and, in some jurisdictions, is subject to revision or suppression. Satellite-derived activity indices are available within days of an overpass and are not subject to the same reporting incentives. For commodity traders tracking iron ore demand, or credit analysts assessing the cash-flow position of a steel producer, an independent physical signal with a consistent methodology across geographies has real value.

Insurance applications are narrower but specific. A business-interruption policy covering a blast furnace typically requires evidence that the insured asset was actually non-operational during the claimed period. Thermal anomaly records provide exactly that: a timestamped, sensor-derived log of whether a furnace was radiating at operational temperatures on any given overpass date. The record is not continuous, and gaps must be acknowledged honestly in any assessment, but it is objective and difficult to dispute.

Satellize applies this method within its broader industrial monitoring capability, running analysis on open Landsat and VIIRS archives and adding commercial tasking where overpass frequency or resolution needs to be improved for a specific client site.

Honest limits of the archive and the index

Landsat's archive extends to 1972 for multispectral data and to 2013 for TIRS thermal data. VIIRS extends to 2012. ASTER scenes over specific industrial sites go back to 2000. That is a meaningful baseline for trend analysis, but it is not a real-time feed. The minimum latency for a processed Landsat scene is roughly 12 to 24 hours after acquisition; VIIRS active-fire and thermal-anomaly products from FIRMS are available within three hours of overpass.

The method works best at large integrated steelworks with multiple blast furnaces, where the thermal footprint is large enough to be unambiguous even at Landsat resolution. Smaller electric-arc mini-mills, particularly those in industrial estates alongside other heat sources, are harder to isolate. Any report derived from this method should state clearly which furnaces are resolvable at which sensor resolution, and which are below the reliable detection threshold.

Typical figures

Spatial resolution (thermal)	100 m (Landsat 8/9 TIRS, resampled to 30 m product); 90 m (ASTER TIR); 375 m (VIIRS I4/I5)
Revisit frequency	~8 days at mid-latitudes (combined Landsat 8+9); daily (VIIRS); on-demand (ASTER, SkySat)
Thermal bands used	MWIR 3.7–4.0 µm (VIIRS I4); TIR 10.6–12.5 µm (Landsat TIRS, ASTER, VIIRS I5)
Sensor saturation threshold	~750 K brightness temperature for Landsat TIRS Band 10; anomaly flag still generated above saturation
Minimum detectable anomaly	Sub-pixel thermal anomalies equivalent to a few hundred square metres of high-temperature source detectable via radiance mixing at Landsat resolution; published fire-detection studies confirm sensitivity to sources well below pixel footprint
Latency from acquisition	12–24 hours (Landsat standard product); ~3 hours (VIIRS via FIRMS near-real-time)
Archive depth	TIRS thermal data from 2013 (Landsat 8 launch); VIIRS from 2012; ASTER TIR from 2000
Cloud penetration	None. Thermal infrared does not penetrate thick cloud; cloud-masking required before analysis
Coverage	Global; Landsat WRS-2 grid covers all land; VIIRS provides daily global coverage
Delivery formats	GeoTIFF anomaly rasters, CSV activity timeseries, GeoJSON site-level status layers, PDF periodic reports

Analytics Satellize can run

Furnace operational status log	Per-pixel brightness temperature anomaly detection against rolling historical baseline; binary active/inactive classification per overpass	Timestamped CSV or GeoJSON log of furnace status per overpass date, with cloud-cover flags and confidence scores
Relative utilisation index	Anomaly detection frequency aggregated over rolling 30- or 90-day windows; normalised against site-specific historical range	Monthly PDF report with utilisation index chart per named facility, benchmarked against prior-year baseline
Multi-site production comparison	Consistent anomaly-detection pipeline applied across a portfolio of named mills; cross-site index normalised to comparable sensor geometry and season	GIS layer and summary table ranking facilities by estimated relative activity for a given period
Shutdown and restart event detection	Change-point detection on the anomaly timeseries; alert triggered when a site transitions from persistent active signal to absence across consecutive cloud-free passes	Near-real-time alert (email or API) within 24–72 hours of confirmed shutdown or restart, with supporting imagery
Business-interruption evidence dossier	Retrospective extraction of all cloud-free thermal passes over an insured site for a claimed outage period; anomaly status assessed for each pass	Structured PDF dossier with per-date sensor imagery, brightness temperature values, and written assessment of operational status; suitable for claims review
Long-run production trend analysis	Anomaly frequency and intensity metrics computed across full TIRS and VIIRS archive (2012/2013 to present) for a named facility or region	Annual trend report with annotated timeseries, correlating satellite-derived activity with publicly reported production data where available

Sources

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.