Satellite Fire Sensors Miss Far More Blazes Than Assumed, New Detection Limits Reveal

Two of the world’s most important fire-watching satellites miss a far larger share of the planet’s blazes than many users of their data realise, according to a new technical study published in Geoscientific Model Development. Mikhail Sofiev and Rostislav Kouznetsov of the Finnish Meteorological Institute have, for the first time, quantified exactly how faint a fire must be before the MODIS and VIIRS instruments can no longer see it, and have built the first proper mathematical bridge between these satellite observations and the numerical models that depend on them. Their results show that even a powerful fire has only a 78 percent chance of being caught by MODIS on a global average, and in some equatorial regions that probability drops below 30 percent.

The work targets Fire Radiative Power, or FRP, a measure of the energy a fire radiates at the moment a satellite passes overhead. FRP products underpin operational global fire information systems, injection-height models that estimate how high smoke plumes rise, and reconstructions of the daily cycle of fire activity. Yet until now, every modelling system that compared predicted fire output with satellite FRP retrievals used what researchers call an identity observation operator, effectively pretending that the modelled quantity and the observed quantity were the same thing. Sofiev and Kouznetsov show that this assumption breaks down badly, because the satellites miss a substantial fraction of fires, a problem known as omission error.

The heart of the study is a simple but powerful relationship: the detection limit of a fire observation is essentially a linear function of the pixel area. A larger pixel averages more background terrain, washing out the thermal contrast that fire detection relies on. For MODIS Collection 6.1, the smallest pixels at the centre of the swath, each about one square kilometre, have a detection limit of 3.7 megawatts for a clear-sky night-time retrieval and 5.7 megawatts by day. At the edge of the swath, where pixels swell to nearly ten square kilometres, the limit climbs to roughly 39 to 42 megawatts, and for a daytime side-view it can reach as high as 40 megawatts. VIIRS, with its much finer 375-metre pixels, is dramatically more sensitive: the smallest fire it reliably detects at night is just 0.5 megawatts.

To pin these numbers down, the researchers processed enormous volumes of raw satellite data. They used MODIS Fire Radiative Power retrievals from the Terra and Aqua satellites covering 2000 to 2023, and VIIRS data from the SUOMI-NPP and NOAA-20 satellites covering 2012 to 2024, all obtained from NASA’s LAADS archive. Roughly 90 million MODIS fire detections fed the statistical analysis. For each pixel size, they built two-dimensional distributions of fire radiative power and fitted the rising edge of each distribution with a shifted logistic function. The midpoint of that sigmoid, where half of fires of a given intensity would be missed, defines the detection limit. The fits were remarkably clean, with correlation coefficients exceeding 0.999 for the linear dependence on pixel area.

The geometry of the scanning instruments turns out to be central to the story. The MODIS swath stretches about 2,063 kilometres across the surface, with pixel area growing from one square kilometre at nadir to 9.66 square kilometres at the swath edge, 55 degrees off vertical. About 40 percent of the swath is covered by pixels smaller than two square kilometres, and only 20 percent by pixels larger than six. Because half of all MODIS fire detections come from the small central pixels, which occupy only about a fifth of the swath, the effective sensitivity of the instrument varies enormously depending on where a fire happens to fall relative to the orbit track, which itself shifts daily on a 16-day repeat cycle.

VIIRS has its own quirks. To keep pixel growth in check, its detectors aggregate neighbouring samples differently across the scan, and readings from the outermost detectors are deliberately discarded to limit data volume, a scheme known as bow-tie deletion. Sofiev and Kouznetsov explicitly accounted for both effects. The payoff is striking: pixels smaller than 0.3 square kilometres occupy 60 percent of the VIIRS swath but deliver about 90 percent of its fire detections. Comparing the two instruments directly, VIIRS is roughly three times more sensitive than MODIS for comparable pixel sizes at night, though about 25 percent less sensitive during the day.

Beyond the detection limit itself, the team constructed two kinds of observation operators. The pixel-level operator converts a model-predicted fire radiative power into what a satellite would actually register, given the pixel size and the day or night conditions. The grid-level operator extends this to the coarse cells used by atmospheric models, adding two further ingredients: the fraction of the cell’s total fire power contributed by its single strongest fire pixel, and the fraction of the cell obscured by clouds, sun glint or missing data. For fully cloudy cells the detection limit is set very high but finite, avoiding numerical problems while effectively zeroing the observed fire output. The operators even include an indicative correction for thin clouds and smoke, which attenuate the 4-micrometre signal the instruments rely on.

Applying these operators produced the study’s most eye-catching results: global maps of fire detection probability. Even during fire seasons, the omission error for low-orbit instruments typically falls within 20 to 40 percent for most fire-prone regions, with VIIRS performing 10 to 20 percent better than MODIS. In equatorial regions with persistent cloud cover, the fraction of missed fires can exceed 60 percent for MODIS and 40 percent for VIIRS. Globally averaged, the probability of detecting a powerful fire is 78 percent for MODIS and 88 percent for VIIRS. The researchers note that fires tend to occur in dry periods when cloud cover is lower than the annual mean, so their fire-season statistics represent bounds on the true omission rate rather than exact values.

The operators also passed a demanding consistency test. The team projected each satellite’s 2020 fire retrievals through the other instrument’s operator, comparing more than 40,000 MODIS Aqua granules that contained fire pixels. Raw totals differed sharply: VIIRS registered 177 terawatts of cumulative fire radiative power against MODIS’s 128 terawatts, a 38 percent gap driven mainly by VIIRS’s lower detection limit. After symmetric application of the observation operators, the processed totals converged to 88 and 106 terawatts, and the remaining 23 percent difference traces to the satellites’ different overpass times, with Aqua observing fires one to two hours later in the day, closer to the estimated peak of fire intensity in the early afternoon.

The implications reach well beyond fire mapping. Atmospheric composition models calibrated against satellite observations can be distorted when a fraction of fires is silently missing: inverse emission estimates may compensate for undetected fires by inflating emission factors for the fires that are seen, a mechanism the authors suggest could partly explain the long-standing factor-of-3.4 discrepancy between top-down and bottom-up estimates of particulate matter emissions from fires. Missing fires also produce appearing and disappearing hot spots as orbits shift, breaking the continuity of smoke plume simulations. The authors propose that a fire prediction model capable of assimilating active-fire products into a continuous model state, using the observation operators developed here, could fill in fires hidden by clouds or hidden below the detection limit. The operator code and its coefficients have been released publicly, and the formulations, though developed for MODIS and VIIRS, can be refitted for other low-orbit and potentially geostationary instruments.

Subject of Research: Detection limits and observation operators for MODIS and VIIRS satellite fire radiative power products

Article Title: Observation operator and detection limits for MODIS and VIIRS Fire Radiative Power products

Article References: Sofiev, M., & Kouznetsov, R. (2026). Observation operator and detection limits for MODIS and VIIRS Fire Radiative Power products. Geoscientific Model Development, 19(20), 10123-10142. https://doi.org/10.5194/gmd-19-10123-2026

Image Credits: AI Generated

DOI: 10.5194/gmd-19-10123-2026

Keywords: MODIS, VIIRS, fire radiative power, satellite remote sensing, detection limit, observation operator, wildfire monitoring, cloud obscuration, omission error, atmospheric modelling, Geoscientific Model Development, Observation