Refinery unit utilisation signals from thermal infrared
Thermal infrared imagery from Landsat 8/9 TIRS and ECOSTRESS reveals the operating state of crude-distillation and catalytic-cracking units through radiant temperature anomalies at process heaters, flare stacks, and cooling towers, providing independent run-rate signals for commodity and credit analysts.
Sensors
- Landsat 8 TIRS: Two thermal bands (Band 10 at 10.6–11.2 µm, Band 11 at 11.5–12.5 µm) at 100 m native resolution, resampled to 30 m in products. 16-day repeat at any given point. Operational since 2013, providing a decade-long archive for baseline construction.
- Landsat 9 TIRS-2: Identical band placement to Landsat 8 TIRS but with improved radiometric calibration and lower detector noise. Launched 2021. When combined with Landsat 8, the effective revisit at mid-latitudes shortens to 8 days, though cloud interference frequently reduces usable acquisitions.
- ECOSTRESS (ISS): Five thermal infrared bands between 8.3 and 12.5 µm at 70 m resolution. Because it flies on the International Space Station rather than a sun-synchronous orbit, its overpass times vary across the day and night, occasionally capturing refinery heat signatures outside the fixed mid-morning window of Landsat. Revisit is irregular, roughly every 1–5 days per site.
- Sentinel-2 MSI (SWIR): Bands 11 and 12 (1.6 µm and 2.2 µm shortwave infrared) at 20 m resolution with a 5-day revisit. SWIR does not measure ambient surface temperature but saturates on very high-temperature targets such as active flare stacks and furnace openings, providing a complementary brightness signal that is particularly useful for confirming flare activity.
What a process heater gives away
A crude-distillation unit running at full capacity forces several hundred thousand barrels of oil per day through a fired heater, raising feed temperatures to roughly 340–370 °C before the fractionating column. That heat does not stay inside the pipe. Refractory walls, stack gases, and radiated energy from the firebox all produce a surface temperature anomaly detectable from orbit. The Stefan-Boltzmann law is unambiguous: radiated power scales with the fourth power of absolute temperature, so even modest differences in operating intensity produce measurable differences in emitted radiance.
Cooling towers are an equally useful signal. A unit running hard circulates more hot process water, and the evaporative plume from the cooling tower grows in both temperature contrast and spatial extent. Landsat TIRS at 100 m native resolution can resolve individual large cooling-tower clusters at major refineries. Smaller units, or refineries with compact footprints, sit at the edge of detectability and require careful multi-date compositing to extract a reliable signal from noise.
Flares as a corroborating channel
Flare stacks are the most thermally intense feature at any refinery, and they are also the most operationally informative. A facility ramping throughput generates more light hydrocarbon off-gas, and flaring intensity tends to track with unit run rates, particularly during start-up and shutdown transitions. Sentinel-2 SWIR bands saturate on active flares, but the saturation footprint itself, measured in pixels, correlates loosely with combustion intensity. For quantitative flare-volume work, VIIRS nighttime fire radiative power is the published standard; for daytime corroboration of unit state, Sentinel-2 SWIR saturation extent is a practical proxy.
One honest limit: flaring also spikes during upsets and emergency venting that have nothing to do with high throughput. A single flare observation is ambiguous. The analytical value comes from combining the flare channel with the process-heater thermal signal and, where available, the cooling-tower plume signature, then reading all three against the historical baseline for that specific facility.
Building a baseline and reading deviations from it
The Landsat archive runs back to 1972 in optical bands and to 2013 in calibrated thermal infrared with TIRS. For any refinery of interest, a multi-year thermal time series can be constructed from cloud-clear acquisitions, typically 15–30 usable scenes per year at mid-latitudes. From that series, analysts derive a facility-specific thermal envelope: the radiance range associated with known operating states, cross-referenced against publicly available refinery maintenance schedules, turnaround announcements, and import/export data where accessible.
Deviations from the envelope are the signal. A sudden drop in process-heater radiance combined with a cooling-tower plume collapse and cessation of flaring is a strong indicator of a planned or unplanned shutdown. A sustained elevation above the historical median suggests the unit is running above its typical utilisation rate. Neither conclusion is certain from thermal data alone, but both sharpen the prior held by a commodity analyst or credit team considerably.
Resolution, revisit, and the cloud problem
Landsat TIRS at 100 m native resolution can distinguish the major thermal sub-units of a large refinery, a CDU heater cluster from a catalytic cracker from a hydrocracker, provided the facility is large enough and the units are spatially separated. At 100 m, a single pixel covers one hectare. A compact European refinery with units packed into a 500 m × 500 m footprint may present as two or three merged thermal blobs rather than individually resolved features. That limits the granularity of inference.
Revisit is the other constraint. Eight days with the combined Landsat 8/9 constellation sounds adequate, but persistent cloud cover over coastal industrial regions, particularly in monsoon climates or high-latitude winters, can reduce usable acquisitions to four or five per quarter. ECOSTRESS partially compensates by acquiring at variable times of day, but its ISS orbit means coverage is not guaranteed on any given date. Analysts should treat 8-day revisit as a best case, not a planning assumption.
What the method cannot do
Thermal infrared inference is a relative, not an absolute, measurement of throughput. It cannot tell you that a refinery processed 180,000 barrels per day last Tuesday. It can tell you that the facility was operating at a thermal intensity consistent with its upper quartile of historical observations, or that it went cold three weeks before the operator disclosed a turnaround. The distinction matters for how the output is used: as a directional signal and anomaly flag, not as a production meter.
Atmospheric correction is a genuine source of uncertainty. Water vapour between the surface and the sensor attenuates thermal emission, and correction algorithms introduce their own error budget, typically ±0.5–1.5 K in land surface temperature retrievals under standard conditions. For high-contrast industrial targets like active furnaces, this uncertainty is small relative to the signal. For cooling-tower plumes operating close to ambient temperature, it matters more. Satellize applies published split-window correction methods using concurrent atmospheric profiles where available, and flags retrievals with high water-vapour uncertainty.
From pixels to an analytic product
The practical output for an energy trader or credit analyst is not a satellite image. It is a time-series chart of thermal anomaly magnitude per facility sub-unit, updated on each cloud-clear acquisition, with a status classification, running, reduced, idle, transitioning, derived from threshold crossings in the historical distribution. Alert notifications can be configured for significant state changes. For a portfolio of refineries across a region, the same pipeline runs in parallel, producing a comparative utilisation index.
Satellize builds these pipelines on open constellation data from Landsat and ECOSTRESS, supplemented by Sentinel-2 SWIR for flare corroboration, and can add commercial thermal tasking for higher revisit on priority sites under client licence. The methodology follows the same radiometric time-series principles applied in the Tonga crop-estimation programme, adapted for industrial thermal targets rather than vegetation indices. Analysts wanting to understand the approach before commissioning a facility study can request a worked example on a publicly known refinery using archived data.
Typical figures
| Thermal spatial resolution | 100 m native (Landsat 8/9 TIRS), 70 m (ECOSTRESS); resampled products at 30 m |
| SWIR spatial resolution | 20 m (Sentinel-2 Bands 11 and 12) |
| Revisit (thermal) | 8 days best case with combined Landsat 8 + 9; irregular 1–5 days with ECOSTRESS on ISS |
| Revisit (SWIR flare channel) | 5 days (Sentinel-2 twin satellites) |
| Spectral bands used | TIR: 10.6–11.2 µm and 11.5–12.5 µm (Landsat TIRS); SWIR: 1.6 µm and 2.2 µm (Sentinel-2) |
| Temperature retrieval uncertainty | ±0.5–1.5 K under standard atmospheric conditions after split-window correction |
| Minimum detectable thermal anomaly | High-contrast industrial targets (furnaces, flares) reliably detectable; cooling-tower plumes near ambient temperature are at the margin of reliable retrieval |
| Archive depth | Landsat TIRS from 2013 (Landsat 8) and 2021 (Landsat 9); Sentinel-2 SWIR from 2015; ECOSTRESS from 2018 |
| Cloud impact | Cloud-contaminated acquisitions are masked; usable scene frequency drops to 4–5 per quarter in persistently cloudy regions |
| Delivery formats | Time-series CSV or JSON feed, GeoTIFF anomaly layers, facility status classification report (PDF or API) |
Analytics Satellize can run
| Facility thermal baseline | Multi-year Landsat TIRS time-series compositing with split-window atmospheric correction; percentile-based operating envelope per facility sub-unit | Historical radiance distribution chart per refinery unit, delivered as PDF report and GeoTIFF archive |
| Unit operating-state classification | Threshold detection against historical thermal envelope; states classified as running, reduced, idle, or transitioning based on radiance percentile crossings | Per-acquisition status label with confidence flag, delivered as structured JSON feed or GIS point layer |
| Shutdown and restart event detection | Change-point detection on thermal time series; corroborated by SWIR flare-brightness collapse or onset | Event alert with timestamp, magnitude, and supporting imagery thumbnail, delivered by API or email notification |
| Flare activity corroboration | Sentinel-2 SWIR Band 12 saturation-extent mapping; compared against concurrent TIRS process-heater anomaly | Flare-active / flare-absent binary flag per acquisition, appended to the thermal status feed |
| Multi-facility utilisation index | Normalised thermal anomaly magnitude ranked across a portfolio of named refineries; relative index updated on each cloud-clear acquisition | Ranked comparative table updated per acquisition cycle, delivered as spreadsheet or dashboard feed |
| Cooling-tower plume analysis | Spatial extent and peak radiance of cooling-tower thermal plumes extracted from TIRS and ECOSTRESS acquisitions; compared against facility baseline | Plume intensity time series per facility, included in quarterly utilisation report |
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.