Russell Buzby

On an afternoon in March this year, an AI-equipped camera in Arizona’s Coconino National Forest noticed a smudge of smoke that no person had reported. The camera had been installed for the utility Arizona Public Service, whose AI alerts are arriving on average about 45 minutes ahead of the first 911 call. Crews reached what became the Diamond Fire while it was still small, and it was contained before it exceeded seven acres, roughly the area of four football fields (Pineda & Peterson, 2026) – Americans will use any measurement except the metric system. Forty-five minutes is not an incremental improvement in fire detection. In wind-driven fuel, it is the difference between a burnt ridge and a burnt town.

Arizona’s camera is one node in a much larger shift. California’s ALERTCalifornia network, run by UC San Diego with CAL FIRE, has grown to 1,263 camera sensors whose AI-screened alerts flow to all 21 of the state’s emergency command centres. In 2025 the system flagged around 3,600 fire incidents, more than half of them before anyone rang 911 (Levinson, 2026). This is no longer a pilot with a press release. It is operational public safety infrastructure with a budget and a roster, catching fires while they are still measured in acres.

The research pipeline behind capabilities like these is, if anything, accelerating. A single week of early-July research digests carried the following. A congestion-aware multi-agent reinforcement learning framework was tested on the road network of Lytton, British Columbia, a town that burned in 2021 (Congestion-aware multi-agent reinforcement learning, 2026). It cut peak evacuation congestion by 74 per cent while routing every vehicle clear of the fire zone, at the cost of a 7.4 per cent increase in average travel distance. Sentinel-2 satellite imagery turns out to carry an anomalous moisture signal that precedes the spring re-emergence of overwintering fires, which accounted for just 1.3 per cent of Canadian ignitions in 2024 but 22.8 per cent of the area burned (Anomalous moisture signal, 2026). Simulations across twelve real satellite constellation designs showed the best-performing configurations can push average detection-information freshness below 70 seconds (Wang, Hu & Gunn, 2026). An open-source Python toolkit, pyflam, now models how fire runs along the surface and through crowns, how embers spot ahead of the front, and how the fire couples with the atmosphere above it, in a package any agency can run without proprietary licensing (pyflam, 2026). And a preprint on what it calls the ignition cascade problem argues that in wildland-urban interface firestorms, simultaneous firebrand showers overwhelm all three traditional defences at once, proposing AI-coordinated neighbourhood fog and mist networks as a fourth layer of engineerable, insurable infrastructure (The ignition cascade problem, 2026). Watching the January 2025 Los Angeles fires, where exactly that simultaneous-ignition dynamic destroyed thousands of structures, it is hard to dismiss the premise as speculative.

Set the list against the state of practice and the pattern is uncomfortable. The science is not the constraint anymore. A fifteen-country comparison of wildfire early warning systems, published the same week, found that maturity tracks national wealth and institutional capacity rather than access to science: the sensing research is largely shared, while the operational systems diverge enormously (Wildfire early warning systems, 2026). Capability now pools where institutions can absorb it, which makes institutional absorption rather than science the binding constraint on how many hectares burn. I made a narrower version of this argument when I wrote that the model is not the bottleneck in Australian fire prediction, where the constraint was data pipelines and operational integration rather than algorithms. This month’s evidence generalises it across the whole stack, from detection to evacuation.

Absorption is genuinely hard, and there is a cautionary tale which shows why. A validation study of an AI fire detection model reported 95 per cent accuracy in controlled testing, falling to 80.2 per cent against real-world footage with atypical smoke and visual noise (Construction and validation, 2026). That 15-point gap is where fire-tech procurement goes wrong. But really, how many times have fire appliances been sent on wild goose chases for smoke sightings after farmers move stock across dusty paddocks, or are spraying? An agency that cannot independently evaluate lab-to-field performance will buy the demonstration, deploy the disappointment, and conclude the technology was hype, when the actual failure was an evaluation capability the agency never built. Integration also means doctrine: who acts on a camera alert that no crew has confirmed, how a control centre weighs an algorithm against a human spotter, when an evacuation model’s routing overrides local judgement. None of that ships with the sensor.

The translational work, research built to be adopted, is where the gap closes, and one of the week’s quieter items is a good Australian example. A design framework for irrigated green firebreaks at the wildland-urban interface, informed by Noosa Shire in Queensland, pairs low-flammability vegetation with recycled urban water and is written in the language of planning schemes rather than journals (Irrigated green firebreaks, 2026). It is a preprint, and untested at scale, but it is aimed at the people who approve subdivisions, which is precisely the audience most fire research never reaches. More of the field’s effort belongs at that boundary, and more of government’s attention belongs on the receiving side of it.

For Australia the timing is pointed. El Niño is declared and the outlooks for the coming season are bad, a subject I write about separately through the lens of prescribed and cultural burning. I have also written about the national bushfire data problem, where the basic picture of what has burned remains fragmented across jurisdictions. Both arguments converge here: a country facing a lengthening fire season, holding world-class fire science, still moves operational capability at the speed of its slowest procurement cycle. California’s cameras took years of sustained funding, standing agreements between a university and a fire agency, and command centres rebuilt around a new information flow. The lesson is institutional patience around a clear operational concept, and it is transferable. What is not transferable is the decade Australia would spend rediscovering it.

The Diamond Fire stopped at seven acres because, some years earlier, a mundane chain held: somebody budgeted for cameras, somebody integrated the alerts into dispatch, and somebody trained crews to trust a machine’s eyes. Fires are decided in their first minutes, and those minutes are decided years earlier by procurement decisions and by operational doctrine that never makes the news. The technology can now see smoke faster than any human being on the continent. It still can’t sign its own purchase order.

References

Anomalous moisture signal in Sentinel-2 imagery precedes overwintering wildfire. (2026, July 2). International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIX-M-1-2026. https://doi.org/10.5194/isprs-archives-xlix-m-1-2026-11-2026

Congestion-aware multi-agent reinforcement learning for wildfire evacuation routing. (2026, July 2). International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIX-M-1-2026. https://doi.org/10.5194/isprs-archives-xlix-m-1-2026-41-2026

Construction and validation of an AI-based fire detection model using fire simulation images. (2026). Journal of the Korean Institute of Fire Science and Engineering. https://doi.org/10.7731/kifse.41c7271c

The ignition cascade problem: How urban wildfire has outgrown its response paradigm, and what to do about it. (2026, June 30). [Preprint]. Figshare. https://doi.org/10.6084/m9.figshare.32832698.v1

Irrigated green firebreaks on wildland-urban interfaces: A conceptual design framework informed by Noosa Shire, Australia. (2026, July 2). [Preprint]. Preprints.org. https://doi.org/10.20944/preprints202607.0168.v1

Levinson, K. (2026, July 1). What California has learned from an AI-enabled early wildfire detection system. Route Fifty. https://www.route-fifty.com/public-safety/2026/07/what-california-has-learned-ai-enabled-early-wildfire-detection-system/414566/

Pineda, D., & Peterson, B. (2026, May 5). States across the wildfire-prone Western US are using AI for early detection.AP via KPBS. https://www.kpbs.org/news/science-technology/2026/05/05/states-across-the-wildfire-prone-western-us-are-using-ai-for-early-detection

pyflam: Open, multiplatform wildfire-behavior modelling. (2026, July 2). [Software]. Zenodo. https://doi.org/10.5281/zenodo.21132875

Wang, Z., Hu, P., & Gunn, G. (2026). Toward LEO satellite network systems for instantaneous detection of environmental changes. arXiv:2605.01243. https://arxiv.org/abs/2605.01243

Wildfire early warning systems: A multisensor and predictive modelling comparison across countries with a Canadian perspective. (2026, July 2). International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIX-M-1-2026. https://doi.org/10.5194/isprs-archives-xlix-m-1-2026-59-2026