In 2025-2026 the global wildfire technology market absorbed roughly $350-400 million in fresh investment. A thermal satellite company (€37M Series B extension), a high-resolution land-surface temperature mission (€37M Series A), a hyperspectral thermal operator ($60M Series B), a hyperspectral constellation backed by a major search company ($13M), and a dedicated camera AI vendor ($44M Series B) are all well capitalised. Yet each of them operates within one of four separate categories: thermal satellites, ground sensors, fire-spread modelling, atmospheric dispersion. The integration layer is empty. No commercial player in the world runs the full end-to-end pipeline — detect a fire, forecast the front, model the smoke plume, compute population exposure, anchor the methodology in peer review. This article explains why the fragmentation persists, what integration delivers operationally, and why this particular niche — with our CJFR 2026 paper as the peer-review anchor and a Ukrainian wartime corpus — is the most defensible position for the next funding wave.
A stack of models, not a product. Integration is what peer reviewers and EU grant panels assess, not an app-store reviewer. Image: ESA, Wikimedia Commons (CC BY-SA 3.0 IGO).
Four separate stacks: the 2026 landscape
Stack 1 — thermal satellite detection. Polar-orbiting VIIRS aboard Suomi-NPP, NOAA-20 and NOAA-21 (NASA FIRMS NRT, 375 m fire-pixel resolution, 3-4 h latency), Sentinel-3 SLSTR with its dedicated F1 channel that does not saturate below 450 K (Xu et al. 2020, Remote Sens. Environ. 248:111947 reported r²=0.9 against airborne FRP), and Landsat 8/9 TIRS (30 m resampled, but with a 16-day revisit usable only for post-fire burn severity via dNBR). Add geostationary GOES-R ABI (2 km, 5-15 min cadence) and the European MTG SEVIRI (10 min cadence, 2 km). On top sit commercial small-satellite constellations with an aspirational 3-10 m ground sampling distance and on-board AI processing (3-5 min latency). Choice here is wide.
Stack 2 — fire spread and evolution. The academic lineage runs from Rothermel 1972 (USDA INT-115, the basis of most operational systems) through FARSITE (Finney 1998, USDA RMRS-RP-4, Huygens elliptical wavelet propagation), FlamMap (Finney 2006, Minimum Travel Time), and ELMFIRE (Lautenberger’s CloudFire, a level-set PDE on a uniform grid that drives the Pyregence operational CONUS forecast), with WindNinja (USFS Missoula, 30-100 m wind downscaling) bolted alongside. The commercial leader is a well-known fire-spread SaaS vendor, which has attracted roughly $84M in cumulative investment and handles 20,000+ incidents a year for CAL FIRE, PG&E and SDG&E. This category already has an established market leader.
Stack 3 — ground sensors. Tower-mounted AI cameras (a dedicated camera AI vendor with $44M Series B, $100M contracted revenue and 30 million acres of coverage), civic networks (ALERTCalifornia at UCSD operates 1,200 cameras), and LoRaWAN IoT sensors in the forest floor (a Berlin-based vendor, €22M raised, sub-60 min detection). Also heavily capitalised.
Stack 4 — atmospheric dispersion and health exposure. The Finnish Meteorological Institute’s SILAM/IS4FIRES (Sofiev group), CAMS GFAS (an ECMWF Kalman-filter assimilation of VIIRS and MODIS FRP at 0.1° daily, 40 species since 2003), and HYSPLIT/BlueSky (USFS AirFire, with a Sonoma Technology front-end). All governmental or academic. Not a single credible commercial actor.
Why the fourth stack is the least served
Smoke dispersion requires three things that rarely live together: numerical weather prediction (NWP) at 1-4 km resolution, Lagrangian chemistry-transport with an explicit source term — FRP and emission factors — and atmospheric chemistry (secondary particle formation: sulphates, nitrates, secondary organic aerosol). The academic community holds the full toolkit (WRF, FLEXPART, CALPUFF, SILAM, CAMS). The commercial side holds almost nothing. The single notable move since 2024 was the October 2023 acquisition of an atmospheric data company by the market-leading fire-spread SaaS vendor, which added dispersion capability to its spread platform; public detail is limited.
The reasons for the fragmentation are structural:
- Computational cost. A fully coupled WRF-Fire run (Mandel, Beezley and Kochanski 2011, Geoscientific Model Development 4:591-610) takes 10-30 minutes on a strong cluster for a 50 km² fire. That does not scale to an operational “every 15 minutes across thousands of fires” target without significant hardware investment.
- Different expertise. NWP is a meteorologist’s domain. Fire spread belongs to a fire scientist. Atmospheric dispersion sits with an atmospheric chemist. Health exposure lives with an epidemiologist. Assembling these disciplines in one team is an organisational problem most startups do not solve.
- Data heterogeneity. VIIRS native resolution is 375 m. WRF runs at 1-4 km. SLSTR F1 is 1 km. A Sentinel-2 burn scar is 10 m. LANDFIRE and FirEUrisk fuel layers are at 30 m and 1 km respectively. Harmonising these layers is a piece of business logic no single team writes without prior experience across all four stacks.
- Absence of a market-side operational demand. A fire agency buys “probability that the containment line holds for the next six hours”. That does not need a dispersion model. An insurer buys “your warehouse is likely to burn within three days”. Still no dispersion. Atmospheric dispersion is needed by public health and civil protection — a market served by government rather than by commerce.
Why this is shifting in 2024-2026
Three external forces are moving demand toward the integrated stack.
The first is the science of smoke mortality. Sofiev et al. 2025 in Nature npj Clean Air and Lancet Planetary Health documented more than 100,000 deaths a year worldwide from wildfire PM2.5. Before that paper, political pressure focused on direct casualties — homes burned, firefighter deaths. After it, attention turned to population exposure within 100-500 km of fires. For the EU this opened a new regulatory vocabulary: smoke exposure as a climate adaptation outcome.
The second is Canada 2023 (18.5 Mha burned, Byrne et al. 2024, Nature 633:835, 647 TgC in emissions — comparable to India’s annual fossil output) and Ukraine 2024 (965,000 ha, more than the EU combined according to JRC). Smoke reached New York, Chicago, the European Arctic and Bulgaria (see the running chronicle in Wildfire Today’s Ukraine archive). The problem of a smoke plume 500-3,000 km from the fire suddenly became an international political subject. CAMS is coarse (10 km); regional adaptive scenarios are needed at national resolution.
The third is Horizon Europe and the EU Mission on Adaptation. Calls for 2026-2028 now explicitly reference cascading multi-hazards (wildfire + drought + heatwave + air quality). The ICARIA project (101093806, 2023-2026) was funded precisely with this grammar. EU grant panels (notably the European Innovation Council) are now looking for teams that can tie detection, modelling, dispersion and exposure together in a single pipeline.
The integration layer: what actually gets connected
The operational WildFiresUA system — our complete stack for Ukraine, built by YourAirTest together with academic partners, runs as follows.
(1) Detection. NASA FIRMS is polled every 10 minutes (VIIRS NOAA-20/21/SNPP NRT as primary, MODIS NRT as secondary). DBSCAN clustering (eps = 1.5 km on the haversine metric with Earth radius 6,371 km, min_samples = 2) collapses multi-pixel detections into event objects. Sentinel-2 MSI (10 m, 5-day revisit) provides post-event verification through dNBR.
(2) Source term for modelling. Each fire object is assigned an FRP (megawatts of fire radiative power), which maps to combustion rate (kg/s) following Wooster et al. 2005 via the 0.368 × 10⁻⁶ kg/J coefficient. Particulate emissions follow Akagi et al. 2011 (Atmos. Chem. Phys. 11:4039) or Andreae 2019 (ACP 19:8523) — emission factors by fuel type (grams of species per kg of fuel burned). For fuel depots and chemical plants these biomass factors do not apply; for those we rely on the SFPE Handbook combined with industrial fire-safety data.
(3) Spread — where the front is going. FARSITE with the Ukrainian 30-metre FBP-code fuel map (O-1b dominance covering 79% of Dnipropetrovsk oblast steppe and a seasonal M-1/M-2 transition through curing between 55% and 85%) for operational and critical-infrastructure scenarios, and ELMFIRE as a cross-check on level-set grid geometry. WindNinja downscales the WRF 1 km wind field to the 30-100 m model that resolves the Dnipro channel, the Carpathian foothills and the steppe corridors.
(4) Atmospheric dispersion. FLEXPART (Lagrangian, from 2023 onward, restricted to wildfire smoke, Stohl et al. 2005, ACP 5:2461-2474) for regional and transboundary plumes — smoke from the Kremenchuk refinery reaching Poltava, smoke from the Kakhovka area crossing into Romania. CALPUFF for radiological scenarios (Chernobyl 2020 — 341 GBq ¹³⁷Cs per Evangeliou et al. 2016, Sci. Rep. 6:26062 and Masson et al. 2022, Atmos. Environ. 291:119402; ZNPP contingency scenarios). HYSPLIT (NOAA Air Resources Laboratory; see NOAA’s dedicated smoke-forecasting guidance) for rapid trajectory queries. All three are driven by WRF-downscaled meteorology.
(5) Population exposure. The plume is projected onto population density (the European High Resolution Settlement Layer or Ukrainian oblast-level statistical data) and converted to aggregate PM2.5 person-hours. The next step applies health impact functions (Burnett et al. 2018, PNAS) to produce attributable DALYs.
(6) Peer-review anchor. Each link of the pipeline is published with methodology and code. The first to appear was our CJFR 2026 paper — a 30-metre fuel map of Ukraine validating the FCCS → CONSUME → FEPS → CALPUFF pipeline against KMDA ground stations (BIAS +2.77 µg/m³ PM2.5, RMSE 48, Pearson r = 0.40). Papers queued next: Earth System Science Data (a data paper on the Ukrainian wartime corpus 2022-2025), Atmos. Chem. Phys. (FLEXPART-WRF coupling for the Kremenchuk 2022 event), Environmental Research Letters (health attribution).
What this delivers operationally
For SESU (the State Emergency Service of Ukraine): after a strike on a fuel depot or refinery, a 48-hour plume forecast with a quantified list of city districts at risk. Not “smoke somewhere” but Darnytskyi district, PM2.5 peaks of 150 µg/m³ between 14:00 and 17:00 tomorrow. This is a level of information product that no single satellite vendor can produce alone.
For municipal authorities: individual alerts for asthma and COPD patients (“your commute today passes through a PM2.5 zone above 100 µg/m³”). The EcoCity network across 80 Ukrainian cities can surface this signal provided a dispersion model runs behind it.
For nuclear power plants and chemical facilities: pre-computed scenarios of the form “if this happens, here is what we recommend”. CZ-170 — a radiological forecast map for the five Ukrainian NPPs — is one such product.
For research and grants: a unique dataset of 20,000+ fires from the 2022-2025 wartime context together with validation measurements, publicly non-reproducible elsewhere. This is the kind of evidence peer reviewers read and remember.
Why the wartime corpus is a genuinely unique scientific resource
JRC EFFIS reported in March 2025 that Ukraine 2024 saw 965k ha burned — more than the entire EU combined. The statistical step change driven by the war is clear: Zibtsev et al. 2024 (Ukrainian Journal of Forest and Wood Science 15(1)) documented 749,500 ha in 2022, 5.2 Mt CO₂, with 68.9% occurring within 60 km of the front line. Researchers elsewhere cannot reproduce this combination of signals: deliberate military ignitions, large-scale damage to energy and chemical infrastructure, radiological contingencies around ZNPP, and dust storms from the drained Kakhovka reservoir bed. This is a unique stress test for any integrated wildfire pipeline.
For Horizon Europe reviewers this uniqueness implies that the validation supplied by the Ukrainian team cannot be obtained any other way. That is a form of defensibility unavailable to US or Mediterranean co-applicants.
Comparison with peers
Every player we examined in the competitive review has a clear category specialism:
- The thermal satellite company sees fire — it does not calculate exposure.
- The market-leading fire-spread SaaS models propagation — it does not deliver smoke to a city block.
- The dedicated camera AI vendor detects early — it does not extrapolate.
- Academic atmospheric dispersion (SILAM, GFAS) has the science — it does not hold operational contracts.
No commercial actor integrates across all four. A handful of European research projects (ICARIA, the FIRE-RES cluster, FirEUrisk) do so academically, but they wind down after a three-year grant. An operational, open-science, peer-reviewed integrator with a national-scale customer does not exist on the market.
For the Ukrainian deep-tech ecosystem
The lesson: do not compete in a category where someone already has $37-100M. Find a wedge between categories where regulatory demand exists, no commercial counterpart is in place, and unique data are available. Atmospheric smoke dispersion combined with fire spread and peer review is exactly such a wedge for us. For other deep-tech teams: CSRD compliance at the agricultural-EU boundary, water-resource security in the post-Kakhovka context, the forestry biocarbon market — analogous wedges exist.
The key is to spend three to six months on competitive intelligence before building an MVP. Look at who is being funded, in what grammar, what the regulator is signalling. The best position is the one no other team can defend because they do not have the data or the academic partnerships required.
Frequently asked questions
Can you be an integrator from scratch, without CJFR?
Yes, but at significantly higher cost and over a much longer horizon. A peer-reviewed publication represents three to four years of data accumulation, validation and work with academic partners. Without it the integration narrative sounds like marketing rather than science. At the $20M+ grant level this is critical.
Why doesn’t the market build dispersion itself?
Because dispersion is paid for by government and public health, not by the private sector. Selling spread-simulation SaaS to CAL FIRE is easier than selling an exposure model to a Centre for Disease Control. Market dynamics push every company into the first three categories.
Aren’t FLEXPART, CALPUFF and HYSPLIT simply outdated open-source tools?
They are the canonical scientific instruments, with decades of validation behind them. A commercial product that “does the same thing, only faster” does not inherit the scientific trust attached to them. Our advantage is that we do not replace FLEXPART — we engineer the pipeline around it, with validated inputs, a peer-reviewed methodology and operationally tuned outputs. That is the difference between engineering and a product launch.
Can the pipeline run on cloud functions instead of a dedicated cluster?
Technically yes. AWS Batch, GCP Cloud Run with GPUs, Azure CycleCloud. For a typical Ukrainian fire season (50-500 active events at peak) this reduces operational cost. For major catastrophes at Canada-2023 scale, dedicated HPC is required.
Ukrainian startup ecosystem: follow TechUkraine and AIN.ua — the two leading outlets covering Ukrainian deep tech, climate tech, and environmental startups.
What to do today
- Check the YourAirTest air quality map for your city — recent PM2.5 readings.
- If the topic resonates, share with colleagues; we monitor referrals via Google Search Console.
- To contribute data (sensor measurements, regional models) reach out via the contact form.
References
Our peer-review paper CJFR 2026: 30-metre Anderson 13 fuel map for Ukraine. Canadian Journal of Forest Research. DOI: 10.1139/cjfr-2025-0035
Finney MA. (1998) FARSITE: Fire Area Simulator — Model Development and Evaluation. USDA Forest Service RMRS-RP-4.
Mandel J, Beezley JD, Kochanski AK. (2011) Coupled atmosphere-wildland fire modeling with WRF 3.3 and SFIRE 2011. Geoscientific Model Development 4:591-610. DOI: 10.5194/gmd-4-591-2011
Stohl A, Forster C, Frank A, Seibert P, Wotawa G. (2005) Technical note: The Lagrangian particle dispersion model FLEXPART version 6.2. Atmospheric Chemistry and Physics 5:2461-2474.
Scire JS et al. (2000) CALPUFF Dispersion Model User’s Guide v5. EPA.
Wooster MJ, Roberts G, Perry GLW, Kaufman YJ. (2005) Retrieval of biomass combustion rates and totals from fire radiative power observations. Journal of Geophysical Research: Atmospheres 110:D24311.
Akagi SK, Yokelson RJ, Wiedinmyer C, et al. (2011) Emission factors for open and domestic biomass burning. Atmospheric Chemistry and Physics 11:4039-4072.
Sofiev M, et al. (2025) Global mortality from wildfire PM2.5 exposure. Lancet Planetary Health.
Jones MW, et al. (2024) Global fire CO₂ emissions from forest fires driven by climate change. Science 386:eadl5889.
Byrne B, et al. (2024) Carbon emissions from the 2023 Canadian wildfires. Nature 633:835.
Zibtsev S, et al. (2024) Assessment of war-induced fires in Ukraine 2022. Ukrainian Journal of Forest and Wood Science 15(1).
European Commission JRC. (2024) EFFIS Advance Report 2024 — Ukraine record-breaking fire season.
NOAA Air Resources Laboratory. HYSPLIT smoke forecasting tools. arl.noaa.gov/hysplit/smoke-forecasting
USDA Forest Service Missoula Fire Sciences Laboratory. WindNinja — diagnostic wind-flow model. ninjastorm.firelab.org/windninja
USDA Forest Service Rocky Mountain Research Station. FlamMap fire-behaviour tool. research.fs.usda.gov
European Innovation Council (EIC) — Horizon Europe grant framework. eic.ec.europa.eu
Wildfire Today — Ukraine news archive. wildfiretoday.com/tag/ukraine