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 in product), 16-day repeat at any given latitude. Night-time acquisitions are essential; daytime solar loading swamps the drum signal. Archive extends to 2013 for Landsat 8.
- ECOSTRESS (ISS-mounted): Five LWIR channels from 8.3 to 12.5 µm at approximately 70 m × 70 m pixel size. Irregular overpass cadence driven by ISS orbital precession gives variable local solar time, enabling some night acquisitions at industrial latitudes. Revisit at a given site can range from daily to several days depending on ISS ground track.
- VIIRS Day-Night Band (DNB): Panchromatic low-light band, 750 m resolution, daily global coverage from Suomi-NPP and NOAA-20. Not a thermal sensor, but the intense visible and near-IR radiance from a hot drum head during venting or steam release can register as an anomaly. Primarily useful as a corroborating signal, not for cycle-phase discrimination.
- VIIRS I-band thermal (I5): 3.74 µm mid-wave infrared band at 375 m resolution, daily revisit. Sensitive to high-temperature point sources. A drum at peak coking temperature (typically 480–510 °C at the furnace outlet, though the drum exterior is cooler) can produce a detectable sub-pixel radiance anomaly. Useful for flagging active-versus-inactive states at the facility level rather than resolving individual drums.
What a coking cycle looks like from orbit
Delayed coking is a thermal cracking process that converts heavy residual oil into lighter products and petroleum coke. The defining operational feature is the drum pair: while one drum fills with hot feed at roughly 480–510 °C furnace outlet temperature, its twin is being cut, cooled with water quench, and prepared for the next fill. The switch happens every 16 to 24 hours depending on feedstock and unit design. That alternation is, in thermal terms, a slow square wave.
From a satellite perspective, the drum in the filling phase presents a hot steel vessel with surface temperatures that can exceed 250 °C on the shell, surrounded by insulation and lagging that partially attenuates the signal. The quenched drum cools rapidly during the water-cutting phase and then equilibrates toward ambient. The temperature contrast between the two states, as seen by a spaceborne LWIR sensor at 70–100 m resolution, is real but modest. A single acquisition tells you very little. A time series tells you a great deal.
Why time-series stacking is the only practical method
Landsat 8/9 TIRS revisits any given site every 16 days. The coking cycle runs every 16 to 24 hours. That mismatch is severe: a single sensor cannot Nyquist-sample the cycle. The approach that works is to stack every available night-time acquisition over a multi-month archive, assign each pixel a brightness temperature, and then look for bimodal distributions or persistent thermal elevation at the drum footprint. A drum that is active for roughly half of all hours will, over enough acquisitions, show a statistically elevated median compared with a drum that is cold or a unit that has been shut down entirely.
ECOSTRESS improves the situation somewhat. Its ISS-driven cadence means it can acquire the same site multiple times within a week during favourable orbital phases, and its 70 m resolution is marginally better than Landsat TIRS. However, ISS orbital precession means coverage at any specific industrial site is irregular and cannot be relied upon for a fixed cadence programme. The practical workflow combines both sensors: Landsat provides the long, consistent archive; ECOSTRESS adds density when available.
Night-time acquisition is non-negotiable. Daytime solar heating raises the apparent temperature of any dark industrial surface, including coke drum shells and surrounding tarmac, by amounts that easily exceed the thermal contrast of the drum cycle itself. Published work on industrial thermal anomaly detection consistently restricts analysis to night passes for exactly this reason.
From drum state to unit count
The analytic objective for commodity intelligence is not the drum temperature itself but the number of coking units running at a given upgrader. A standard delayed coker consists of one furnace feeding two drums. A large upgrader, such as those operating in the Canadian oil sands or at major US Gulf Coast refineries, may have three to six coking units. Each unit has a distinct spatial footprint that can be identified in high-resolution optical imagery and then used as a region of interest for the thermal time series.
Unit-count estimation works by classifying each drum pair as active or idle in each night-time acquisition, then aggregating across the archive to produce a utilisation fraction. A unit that shows elevated thermal signal in more than roughly half of night acquisitions over a 90-day window is likely running. One that is consistently cold has probably been shut down for maintenance or feedstock reasons. The resulting output is a facility-level operating rate, expressed as a fraction of nameplate coking capacity, updated as new acquisitions arrive.
Honest limits: what this method cannot do
Cloud cover is the primary data gap. Thermal infrared does not penetrate cloud. Upgraders in northern Alberta operate through winters where multi-week cloud-persistent periods are common. During those windows, no thermal acquisition is usable, and the time-series estimate carries increasing uncertainty. There is no workaround within the LWIR approach; SAR-based proxies (construction activity, flare detection) can partially fill the gap but address different signals.
Spatial resolution is a genuine constraint at smaller facilities. At 100 m TIRS resolution, two drum pairs separated by less than 150 m may not be individually resolved. The method works best at large upgraders where units are spread across a wide footprint. At compact refineries, drum-level discrimination may collapse to facility-level status only.
The 16-to-24-hour cycle period, combined with a 16-day Landsat revisit, means the phase of the cycle at acquisition time is essentially random. You cannot determine whether a drum is in the early, middle, or late filling phase from a single pass. The statistical approach described above averages over this uncertainty, but it requires at minimum 20 to 30 usable night acquisitions before the utilisation estimate stabilises. That implies a minimum archive window of six to twelve months for Landsat-only analysis.
Finally, the method estimates operating state, not throughput volume. A drum running at 60 % of design feed rate looks thermally similar to one running at 100 %. Feed-rate inference requires additional proxies from other pages in this library.
Putting the signal in context for commodity analysis
Delayed coking capacity is a critical variable in heavy crude upgrading economics. When coker units go offline, the upgrader cannot process bitumen or vacuum residue, and the facility's effective crude throughput drops regardless of how the upstream units are running. Tracking coker operating rates at the major upgraders in Alberta, the US Midwest, and the US Gulf Coast provides a leading indicator of heavy crude demand and petroleum coke supply that is independent of company disclosures.
Satellize applies this thermal time-series methodology as part of its refinery analytics work, running on open Landsat and ECOSTRESS archives supplemented by commercial tasking where revisit density needs to increase. If you are evaluating this approach for a specific upgrader or refinery cluster, the right first step is a feasibility assessment against the existing archive for that site, which will show how many usable night acquisitions exist and whether the drum footprints are individually resolvable at current sensor resolution.
Typical figures
| Spatial resolution (Landsat 8/9 TIRS) | 100 m native, resampled to 30 m in L2 product; individual drum pairs resolvable only at large upgraders with wide unit spacing |
| Spatial resolution (ECOSTRESS) | Approximately 70 m × 70 m; marginal improvement over Landsat TIRS for drum-pair discrimination |
| Revisit cadence (Landsat 8 + 9 combined) | 8 days at equator, 16 days per satellite; usable night acquisitions depend on cloud cover and solar geometry |
| Revisit cadence (ECOSTRESS) | Irregular, driven by ISS orbital precession; can range from daily to several days at a given site |
| Spectral bands used | LWIR 10.9 µm and 12.0 µm (Landsat TIRS); 8.3–12.5 µm multi-channel (ECOSTRESS); 3.74 µm MWIR (VIIRS I5 for corroboration) |
| Minimum archive window for stable estimate | 6 to 12 months of night acquisitions for Landsat-only; shorter with ECOSTRESS supplementation |
| Minimum detectable thermal contrast | Published TIRS noise-equivalent temperature difference approximately 0.4 K; drum-versus-ambient contrast at shell surface typically well above this threshold, but sub-pixel mixing reduces effective sensitivity |
| Cloud limitation | LWIR blocked by cloud; persistent overcast periods in northern Alberta can produce data gaps of weeks; no thermal workaround |
| Archive depth | Landsat 8 from February 2013; Landsat 9 from October 2021; ECOSTRESS from July 2018 |
| Deliverable latency | Landsat L2 products available within 12 hours of acquisition; ECOSTRESS products typically within 24 hours; analytic update cadence depends on acquisition frequency |
Analytics Satellize can run
| Drum-pair operating state classification | Brightness temperature extraction at known drum footprint coordinates, bimodal distribution fitting across night-time archive, threshold classification into active/idle states | Per-drum-pair status table updated on each new night acquisition, delivered as a structured data feed or GIS point layer |
| Facility-level coker utilisation fraction | Aggregation of drum-pair state classifications across all units at a facility, expressed as fraction of nameplate coking capacity assessed as active over a rolling 30- or 90-day window | Monthly or quarterly utilisation report per facility, with time-series chart and confidence interval reflecting cloud-gap frequency |
| Maintenance event detection | Persistent cold-drum signal over multiple consecutive acquisitions, cross-referenced against flare activity and optical activity proxies to distinguish planned turnaround from unplanned outage | Outage alert with estimated start date, duration, and affected unit count; delivered as a structured alert or PDF summary |
| Multi-site upgrader operating rate index | Aggregation of facility-level utilisation fractions across a defined set of upgraders (e.g., Alberta oil sands complex or US Gulf Coast heavy crude belt), weighted by nameplate coking capacity | Weekly index value with site-by-site breakdown, delivered as a spreadsheet or API feed for integration into commodity models |
| Thermal periodicity characterisation | Where acquisition density is sufficient (ECOSTRESS-dense periods), autocorrelation analysis of the brightness temperature time series to estimate cycle period and confirm it falls within the published 16–24-hour range | Site-specific cycle period estimate with confidence bounds, included in feasibility assessment report |
| Archive feasibility assessment | Retrospective query of Landsat and ECOSTRESS archives for a target site, counting usable cloud-free night acquisitions and assessing drum-footprint resolvability at current sensor resolution | One-page site feasibility report with acquisition count, cloud-gap histogram, and resolution adequacy assessment |
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.