Active wildfire front detection from thermal infrared

Mid-wave and thermal infrared sensors detect actively burning fire fronts by measuring radiance that saturates standard land channels, enabling near-real-time alerts. Latency, resolution floors and cloud cover set hard limits on what any system can reliably report.

Sensors

VIIRS (SNPP / NOAA-20): The I-band fire channel (I4, 3.74 µm) resolves fire pixels at 375 m; the M-band product runs at 750 m. Together the two satellites provide up to four overpasses per day at mid-latitudes. The VNP14 active fire product is the operational standard for near-real-time detection.
MODIS Terra / Aqua: Bands 21 and 22 (3.959 µm) and Band 31 (11 µm) underpin the MOD14/MYD14 fire products at 1 km resolution. Two overpasses per day per satellite. Sensitivity is lower than VIIRS for small or low-intensity fires; the 1 km pixel can mask a fire occupying only a fraction of its area.
Sentinel-3 SLSTR: The Sea and Land Surface Temperature Radiometer carries dedicated fire channels at 3.74 µm and 10.85 µm with a 1 km nadir pixel and a roughly 27-hour revisit at the equator (shorter at higher latitudes). Fire Radiative Power (FRP) is a standard Level 2 product. Useful for cross-validation and for tracking large, persistent fronts.
Landsat 8 / 9 TIRS: TIRS Band 10 (10.6–11.19 µm) at 100 m resolution (resampled to 30 m in products) provides the finest thermal pixel available on a free, systematic archive. The 16-day revisit makes it unsuitable for real-time alerting but valuable for cross-validating VIIRS detections and for post-event reconstruction of fire progression.

Why a burning tree saturates a land-surface sensor

Standard optical and thermal sensors are calibrated for Earth's surface temperature range, roughly 200–330 K. An active fire front burns at 600–1200 K. At those temperatures, mid-wave infrared radiance (around 3.9 µm) increases by orders of magnitude relative to background, which is why MODIS Band 21 was given a separate, attenuated detector specifically to avoid saturation. VIIRS I4 uses a similar design. The physics is straightforward Planck-curve arithmetic: at 3.9 µm, a 1000 K source emits roughly 80,000 times more radiance than a 300 K surface.

Fire Radiative Power (FRP), measured in megawatts per pixel, is the primary quantitative output. It is derived from the radiance excess above the background, using the method published by Wooster et al. (2003) and subsequently embedded in both the MODIS and VIIRS operational processing chains. FRP is proportional to the rate at which biomass is consumed, which makes it useful not only for fire mapping but also for smoke-emission estimation.

What the operational products actually deliver

The VIIRS VNP14 product, produced jointly by NASA and NOAA, flags individual 375 m pixels as fire, cloud, water or clear land. Each flagged pixel carries a confidence level (low, nominal, high) and an FRP estimate in MW. NASA's Fire Information for Resource Management System (FIRMS) distributes these detections globally with a latency that is typically 3 hours from overpass but can reach 6 hours during high-traffic periods. That latency is not a software problem; it is the time required for the satellite to downlink, for ground stations to process, and for the product to propagate through the distribution chain.

MODIS MOD14/MYD14 runs at 1 km. At that resolution, a fire occupying 10 % of a pixel may go undetected if its temperature is not extreme. The practical detection threshold for VIIRS at 375 m is roughly 4–5 MW under good conditions; for MODIS at 1 km it is considerably higher, perhaps 50 MW, though both thresholds vary with background temperature and viewing angle. Neither system reliably detects smouldering fires or fires beneath dense smoke, because smoke attenuates mid-wave infrared signal and cloud completely blocks it.

Sentinel-3 SLSTR's FRP product uses a similar contextual algorithm. Its value lies partly in its independent calibration and partly in the fact that two Sentinel-3 satellites (A and B) together reduce the revisit gap. For large fires burning for multiple days, SLSTR provides a useful independent time series.

Latency is the operational constraint nobody advertises

For a fire crew or an emergency operations centre, the difference between a 3-hour-old detection and a 6-hour-old detection can be the difference between a useful tactical picture and a historical one. Polar-orbiting sensors have fixed overpass windows. At 50° latitude, VIIRS SNPP and NOAA-20 together may offer four overpasses per day, but those overpasses are clustered: two in the morning and two in the afternoon or evening. A fire that ignites at noon and is assessed at the next overpass at 17:00 has had five hours to grow.

Geostationary sensors (GOES-16/17 ABI, Meteosat SEVIRI) provide 10-minute refresh rates over their fixed footprints and are the standard for real-time fire monitoring in the Americas and Europe respectively. Their spatial resolution is coarser (2 km for ABI fire channels), but for tracking a moving front in near-real-time, a 2 km detection every 10 minutes is operationally superior to a 375 m detection every 6 hours. Any serious fire-monitoring workflow uses both: geostationary for temporal density, VIIRS for spatial precision.

Landsat's role: precision where it matters, not speed

Landsat 8 and 9 TIRS, at 100 m native thermal resolution, can resolve individual fire lines and distinguish multiple burning fronts within a single VIIRS pixel. This matters for post-event reconstruction: overlaying a TIRS acquisition on VIIRS detections from the same day can reveal whether a VIIRS 'high confidence' pixel was a single intense fire or a cluster of smaller ones. It also matters for cross-sensor calibration studies, which is why Landsat TIRS data appears regularly in the peer-reviewed fire-detection literature.

The 16-day revisit is the hard constraint. Landsat is not a fire-alerting tool. It is a precision instrument for answering specific questions after the fact, or for the occasional coincidence when a Landsat overpass happens to catch an active front. When that coincidence occurs, the imagery is exceptional.

Honest limits of the thermal infrared approach

Cloud cover is the most significant gap. Thick pyrocumulus cloud, which large fires generate themselves, can completely obscure the fire below it from any passive optical or infrared sensor. Smoke attenuates mid-wave infrared less severely than visible wavelengths, but dense smoke still reduces apparent FRP. Neither VIIRS nor MODIS can see through cloud; the confidence flags in VNP14 account for cloud proximity but cannot recover obscured pixels.

Small fires are systematically under-detected. Agricultural burns covering less than a hectare, or early-stage fires in dense canopy, frequently fall below the detection threshold. The 375 m VIIRS pixel is a meaningful improvement over 1 km MODIS, but a fire must still emit enough radiance to stand out from background within that pixel. Night-time acquisitions improve sensitivity because background radiance is lower, which is why night-time VIIRS passes are particularly valuable for detecting low-intensity fires.

Satellize runs VIIRS and FIRMS data through its analytics pipeline for clients requiring automated alerting and FRP time-series reporting. The Tonga crop-estimation programme is a separate engagement, but the same sensor stack and processing infrastructure underlies both.

Building a detection workflow that accounts for the gaps

A defensible operational workflow fuses at least two independent detection streams. The standard approach pairs a geostationary product (GOES ABI or SEVIRI, depending on geography) with VIIRS VNP14 for spatial confirmation. SLSTR FRP provides a third, independent estimate for large events. Landsat TIRS enters the workflow for post-event analysis or for calibrating local detection thresholds.

Confidence filtering matters. Raw FIRMS feeds contain low-confidence detections that include industrial heat sources, gas flares and sun-glint artefacts. Filtering to 'nominal' or 'high' confidence, then applying a spatial clustering step to remove isolated single-pixel detections, substantially reduces false positives without materially reducing true fire detections. The resulting alert feed is smaller, more actionable, and far less likely to exhaust the attention of an operations team.

Typical figures

Best spatial resolution (fire pixel)	375 m (VIIRS I-band); 1 km (MODIS, Sentinel-3 SLSTR); 100 m native thermal (Landsat TIRS)
Revisit frequency	Up to 4× per day at mid-latitudes (VIIRS SNPP + NOAA-20 combined); 1–2× per day per satellite (MODIS); ~27 h equatorial (Sentinel-3 SLSTR); 16 days (Landsat 8/9)
Typical product latency (polar orbit)	3–6 hours from overpass to FIRMS distribution for VIIRS; similar for MODIS
Key spectral bands	Mid-wave IR: ~3.7–4.0 µm (VIIRS I4/M13, MODIS Bands 21/22, SLSTR S6); Thermal IR: ~10.8–12 µm (MODIS Band 31, SLSTR S8/S9, Landsat TIRS Band 10)
Approximate minimum detectable FRP	~4–5 MW per 375 m pixel (VIIRS, night, good conditions); ~50 MW per 1 km pixel (MODIS); both thresholds degrade under smoke or high background temperature
Cloud penetration	None. Passive thermal IR cannot see through cloud or dense pyrocumulus
Global coverage	Daily global coverage (VIIRS, MODIS); regional geostationary gap-fill at 10-minute refresh (GOES, SEVIRI, not covered here)
Archive depth	MODIS from 2000; VIIRS SNPP from 2012; Landsat thermal from 2013 (L8) and 2021 (L9); Sentinel-3 SLSTR from 2016
Primary quantitative output	Fire Radiative Power (MW per pixel), fire pixel confidence class, fire pixel count

Analytics Satellize can run

Near-real-time fire front alert	Contextual threshold algorithm (Giglio et al. / VNP14 method) applied to VIIRS I4 radiance; confidence filtering and spatial clustering to suppress false positives	GeoJSON alert feed or webhook notification, updated per overpass, with pixel-level confidence and FRP estimate
Fire Radiative Power time series	FRP aggregation from VNP14 and MOD14/MYD14 over a defined area of interest; gap-flagging for cloud-obscured overpasses	CSV or dashboard chart showing FRP (MW) per overpass over the incident duration, with cloud-gap annotations
Multi-sensor fire front comparison	Spatial overlay of VIIRS 375 m, SLSTR 1 km and (where available) Landsat TIRS detections for the same 24-hour window; disagreement flagged for analyst review	GIS layer (GeoPackage or Shapefile) with per-pixel sensor agreement score and FRP from each source
Fire progression map	Sequential VIIRS passes composited to show spatial advance of fire perimeter over time; perimeter extracted by convex hull or alpha-shape on high-confidence pixels	Animated GIF or time-stamped GeoJSON sequence showing fire front advance at overpass resolution
False-positive filtered operational alert	Industrial heat source and persistent anomaly masking using multi-day VIIRS baseline; gas flare and smelter locations cross-referenced against known facility databases	Cleaned alert feed suitable for direct ingestion by emergency operations centre GIS, with suppressed industrial detections logged separately
Smoke-obscuration risk flag	VIIRS cloud and aerosol optical depth flags combined with active fire pixel density to estimate probability that adjacent undetected pixels contain fire	Per-overpass uncertainty layer appended to the alert GeoJSON, flagging zones where detection confidence is degraded by smoke or cloud

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.