Industrial cooling tower water vapour plume mapping for facility activity inference
Wet cooling towers at power stations emit visible water vapour plumes whose size and persistence betray thermal load and generation output. High-resolution optical imagery, normalised against meteorological reanalysis, turns those plumes into an activity signal even when thermal infrared retrievals fail under cloud.
Sensors
- Planet SuperDove: 3 m native resolution, 8 spectral bands including a narrow NIR band well suited to condensation contrast against land backgrounds. Revisit of roughly 1 to 2 days at mid-latitudes from the 200-satellite flock, enabling near-daily plume state capture. Tasking is on-demand under commercial licence.
- Sentinel-2 MSI: 10 m resolution in visible and NIR bands, 5-day revisit at the equator (2 to 3 days at higher latitudes with both satellites). Free and open archive back to 2015. The 60 m coastal aerosol band adds atmospheric scattering context. Cloud masking is the primary operational constraint.
- Landsat 8/9 OLI-TIRS: 30 m multispectral resolution with a 16-day revisit per satellite (8 days combined). The thermal infrared bands (100 m, resampled to 30 m) allow simultaneous plume optical and thermal characterisation in the same acquisition, directly linking condensation extent to surface heat signature.
- Sentinel-3 OLCI: 300 m resolution, 1 to 2 day revisit. Too coarse to resolve individual tower plumes at most facilities, but useful for regional atmospheric humidity context and cross-validation of plume direction against meteorological reanalysis fields.
What a cooling tower plume actually encodes
A wet cooling tower works by evaporating a fraction of its circulating water into the airstream. The visible white plume is condensed water vapour: it forms when the warm, saturated exhaust air mixes with cooler ambient air and crosses the dew point. The plume's length, width and persistence are not arbitrary. They are functions of the tower's thermal duty, ambient temperature, relative humidity and wind speed.
That dependency is the measurement opportunity. At constant meteorological conditions, a longer and denser plume means higher thermal load, which means higher generation output. When meteorological conditions vary, reanalysis products such as ERA5 (from ECMWF, available at roughly 31 km grid spacing and hourly temporal resolution) provide the normalisation layer needed to isolate the facility-driven signal from the atmospheric background. The result is a proxy for generation state: running hard, idling, or off.
Why optical imagery works when thermal does not
Thermal infrared retrievals from Landsat TIRS or similar sensors require a clear atmospheric column. Under overcast conditions, cloud emission masks the surface signal entirely. Water vapour plumes, by contrast, are themselves visible in reflected shortwave bands: they scatter sunlight strongly in the 400 to 900 nm range and appear as high-reflectance features against darker land backgrounds.
Planet SuperDove's 3 m resolution is sufficient to distinguish individual tower plumes at most large facilities, where cooling tower spacing runs to tens of metres. Sentinel-2 at 10 m resolves plume extent reliably once the plume length exceeds roughly 50 to 100 m, which corresponds to moderate or higher load conditions. Below that length, plumes collapse into sub-pixel features and the optical method loses sensitivity. This is the honest lower bound of the technique: it cannot confirm that a facility is off, only that its plume is below the detection floor.
On partly cloudy days, the two methods are complementary. Where cloud gaps permit thermal retrieval, the TIR signal anchors the absolute heat flux estimate. Where cloud fills the gap, the optical plume extent provides a directional activity indicator. Neither method alone is sufficient for continuous monitoring; the combination is more informative than either in isolation.
Plume segmentation: from pixel to measurement
Automated plume delineation uses spectral thresholding on the visible and NIR bands, exploiting the high reflectance of condensed water droplets relative to most land covers. A normalised difference approach, comparing a visible band against a shortwave infrared band where liquid water absorbs strongly, sharpens the plume boundary against mixed backgrounds including thin cloud. Sentinel-2 band 3 (green, 560 nm) and band 11 (SWIR, 1610 nm) form a useful pair for this purpose, as published studies on cloud and fog detection have demonstrated.
Once the plume polygon is extracted, its projected area and downwind length are the primary measurements. Wind direction and speed from ERA5 or MERRA-2 reanalysis are used to rotate the plume into a facility-relative coordinate frame, removing the directional ambiguity that arises when the plume blows back over the facility itself. Atmospheric stability class, derived from reanalysis temperature profiles, further conditions the expected plume rise and spread, providing the normalisation denominator for cross-date comparison.
Plume shadow is a secondary observable. On high-sun-angle acquisitions, a dense plume casts a measurable shadow on the ground surface downwind of the towers. Shadow length is geometrically related to plume height, which in turn depends on buoyancy flux and therefore thermal output. This is a noisier signal than plume area, but it is independent of the reflectance thresholding approach and provides a useful cross-check.
Meteorological normalisation: the step most analyses skip
Raw plume area is almost meaningless without normalisation. A facility running at half capacity on a cold, humid winter morning may produce a larger plume than the same facility at full capacity on a dry summer afternoon. The physics is straightforward: lower ambient temperature reduces the mixing needed to reach saturation, and higher ambient humidity means less evaporative cooling is required before condensation begins.
The normalisation procedure regresses observed plume area against ERA5 2 m temperature, 2 m dewpoint depression and 10 m wind speed at the facility location and acquisition time. The residual from this regression is the meteorologically adjusted activity index. Across a time series of acquisitions, this index tracks generation state with considerably less variance than raw plume area. The method does not produce a megawatt figure directly; it produces a relative activity signal that can be calibrated against known generation records where those are publicly available, or used in a purely relative sense to flag changes from baseline.
Limits, confounders and what the data cannot tell you
Several confounders deserve honest acknowledgement. Forced-draught and induced-draught dry cooling towers produce no visible plume regardless of load; the method applies only to wet towers. Some facilities operate hybrid wet-dry systems whose plume behaviour is load-dependent in a more complex way. Natural-draught hyperbolic towers, the large concrete structures common at nuclear and older coal stations, produce more persistent plumes than mechanical-draught towers at equivalent loads, so facility type must be characterised before cross-facility comparison.
Cloud cover is the dominant data-gap driver. At mid-latitude sites with frequent overcast, Planet SuperDove's near-daily revisit still yields clear acquisitions only on the order of 10 to 20 days per month in winter. Sentinel-2's 5-day revisit worsens this. A monitoring system that requires continuous daily signals must accept that optical plume mapping will have gaps, and should plan for gap-filling from thermal anomaly products on days when partial clearing permits TIR retrieval.
Finally, plume absence does not confirm shutdown. A facility may be operating in dry-cooling mode, or the plume may be below the detection floor, or cloud may be obscuring it. Confident shutdown attribution requires corroboration from a second signal, such as absence of a nighttime lighting signature or a flat thermal background on a clear-sky acquisition. Satellize's multi-signal fusion approach for coal plant activity, described in the sibling page on that topic, addresses this directly.
From imagery to intelligence: what the analytic pipeline delivers
The practical output of a plume-mapping programme is a time-stamped activity index at each monitored facility, updated on every clear-sky acquisition. For a portfolio of, say, twenty power stations across a country, that index provides an independent check on reported generation figures and an early-warning signal for unannounced outages or capacity additions.
Satellize runs this kind of analytics on open constellations and adds commercial Planet tasking where revisit frequency matters. The Tonga crop-estimation programme demonstrated that meteorological normalisation of satellite signals is operationally tractable at national scale; the same normalisation logic applies here, applied to a different physical process. Clients who want to trial the method can request a single-facility retrospective analysis over a defined date range using existing Sentinel-2 archive, which requires no new tasking and can be scoped and delivered within weeks.
Typical figures
| Best spatial resolution (optical) | 3 m (Planet SuperDove); 10 m (Sentinel-2 MSI visible/NIR) |
| Minimum detectable plume length | Approximately 50 to 100 m at 10 m resolution; finer at 3 m, subject to contrast with background |
| Revisit frequency | 1 to 2 days (Planet SuperDove, commercial); 2 to 5 days (Sentinel-2, free); 8 days combined (Landsat 8+9) |
| Effective clear-sky acquisition rate | Highly site-dependent; typically 10 to 20 usable acquisitions per month at mid-latitudes in winter |
| Spectral bands used | Visible (400 to 700 nm), NIR (700 to 900 nm), SWIR (1600 nm) for plume segmentation; TIR (10 to 12 µm) for thermal corroboration |
| Meteorological normalisation source | ERA5 reanalysis (ECMWF): 31 km grid, hourly; or MERRA-2 (NASA): 50 km grid, hourly |
| Archive depth | Sentinel-2: from 2015; Landsat 8: from 2013; Landsat 9: from 2021; Planet SuperDove: from 2021 (commercial) |
| Latency (operational monitoring) | Sentinel-2 Level-2A typically available within 3 to 5 hours of acquisition; Planet tasking delivery within 24 hours |
| Delivery formats | GeoTIFF plume polygons, CSV activity index time series, PDF facility reports |
| Facility types supported | Wet natural-draught and mechanical-draught cooling towers at nuclear, coal and gas stations; dry towers are outside scope |
Analytics Satellize can run
| Plume area time series per facility | Spectral thresholding on Sentinel-2 or Planet imagery using visible/SWIR normalised difference; cloud-masked using scene classification layer | CSV time series and GeoTIFF polygon archive per facility, updated on each clear-sky pass |
| Meteorologically adjusted activity index | Regression of raw plume area against ERA5 2 m temperature, dewpoint depression and wind speed at acquisition time; residual treated as load proxy | Monthly activity index report with confidence bands, flagging anomalous deviations from facility baseline |
| Plume shadow height estimate | Solar geometry and shadow length measurement from high-sun-angle Planet acquisitions; buoyancy flux inference from plume rise | Supplementary plume height layer in GeoTIFF, available on acquisitions with sun elevation above 40 degrees |
| Facility operational state classification | Rule-based classifier combining plume area index, thermal anomaly presence (Landsat TIRS) and nighttime lighting; states: active-high, active-low, indeterminate, off | Per-acquisition state label in CSV feed; alert triggered on transition from active to indeterminate or off |
| Multi-facility portfolio dashboard | Aggregation of per-facility time series across a defined asset list; cross-facility normalisation to enable relative load comparison | Interactive GIS layer and PDF summary covering all monitored facilities, delivered monthly or on request |
| Retrospective baseline characterisation | Analysis of Sentinel-2 archive (2015 to present) to establish seasonal plume-area norms and meteorological sensitivity coefficients per facility | Facility baseline report with seasonal profiles, used to calibrate ongoing operational monitoring |
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.