The race to teach AI to fight fires

Artificial intelligence is emerging as a critical tool in the escalating fight against wildfires.

As climate change fuels more intense and erratic fire seasons, scientists and startups are racing to build artificial intelligence systems that can detect, predict and even model fire behavior before first responders arrive. What was once a slow-moving hazard has become the site of a fast-paced, data-driven battle, and AI is increasingly at the center of it.

“We’re not just detecting smoke,” Sonia Kastner, CEO of wildfire detection startup Pano AI, told IBM Think in an interview. “We’re giving responders the kind of early head start that can prevent a disaster.”

Pano AI recently closed a USD 44 million Series B round, bringing its total funding to USD 89 million, to scale its wildfire detection infrastructure and sharpen its real‑time alerting tools. The system identifies wildfire smoke using AI and a dataset of images, sending verified alerts to fire agencies minutes after blazes start.

This type of early detection is becoming a central goal for public safety officials and scientists alike, who say they need faster and more accurate information. Wildland firefighters are now expected to make split-second decisions with partial information, often in rugged or remote terrain. Fortunately, several researchers and companies are now attempting to utilize AI to enhance this view, combining data from cameras, satellites and models to help identify small ignitions before they escalate into catastrophes.

“What’s been most striking is just how under-resourced wildland firefighters are,” Kastner said. “They’re doing high-stakes work without the tools they need. The takeaway is clear: The people on the front lines are ready. What they need are tools that meet the urgency of the threat.”

Firefighting crews may still be outfitted with chain saws and radios at the front lines, but the hope is that help might eventually arrive from a very different kind of toolkit. At the University of Southern California, in a quiet lab far removed from any fire line, Assad Oberai, the Hughes Professor and Professor of Aerospace and Mechanical Engineering, is trying to give responders something they’ve never really had before—foresight. He is developing AI models designed to predict fire behavior in places where traditional, physics-based models tend to struggle, such as forests and densely populated suburban zones.

“We have some reasonably good models that predict wildfire spread in natural settings,” Oberai told IBM Think in an interview. “But in places like the hills around Los Angeles, where natural vegetation meets homes, the physics is extremely complicated. That’s where AI could offer a path forward.”

According to Oberai, one of the biggest technical challenges in using AI for wildfires is speed. Models have to deliver fast, actionable results that are easy for first responders to interpret under stress. They also have to account for uncertainty, such as variations in weather patterns, vegetation and terrain, while still producing functional, reliable scenarios.

“There’s no single answer for what a fire will do next,” Oberai said. “But a good system can show you the range of possibilities and help you understand the risk in real time.”

The idea isn’t to predict the future with certainty, but to narrow the unknowns, giving first responders something closer to a weather forecast than a coin toss. That’s the promise behind systems like Pano AI’s, which pair machine learning with an old firefighting standby: a good view.

Mounted on towers and remote hillsides, Pano AI’s cameras sweep the horizon every 60 seconds, scanning for the telltale smear of wildfire smoke. When something looks suspicious, the images are sent not to an algorithm alone, but also to a team of human analysts, who sort fire from fog in real-time. If they confirm a fire, the system creates a full alert, complete with GPS coordinates, live video, wind speed and temperature, and sends it directly to fire departments and emergency managers.

The goal is not just to be fast, but to be trustworthy. False alarms waste time and resources, especially in regions where fire crews are already stretched. Kastner said Pano AI's system is trained to distinguish wildfire smoke from fog, steam, dust and other visual noise. The alerts are designed to be both early and accurate.

“Speed on its own isn’t helpful unless it leads to the right response,” Kastner said. “We’re focused on delivering alerts that are actionable, tuned to the local terrain and conditions.”

Unlike early smoke detection tools, which relied on line-of-sight observers or weather forecasts, today’s systems are built on vast amounts of imagery, cloud computing and machine learning. Pano’s own training dataset includes image sequences from four fire seasons, ranging from arid outback regions in Australia to the snowy forests of British Columbia. That variety, Kastner said, helps the system learn what smoke looks like under different conditions and across various ecosystems.

But the fire itself is only part of the equation. AI is also being applied to estimate where fires are most likely to ignite, using terrain, fuel levels and real-time environmental data. In some cases, drone and satellite imagery are used to map fuel sources in advance. In others, the goal is to monitor known ignition points, such as power lines or dry lightning corridors.

Still, Oberai emphasized that many of the most hazardous fires occur in difficult-to-model transitional zones, like wild-urban interfaces.

“These are precisely the locations where fires can be deadliest. They involve a combination of trees, brush, roads, buildings, fences, decks and fuel tanks. You have natural and human-made structures side by side. That makes it very difficult to simulate with traditional tools.”

AI can help overcome that complexity by learning from data, rather than theory,  Oberai said. But it must be trained carefully and evaluated against real-world events. He is cautious about the limits of any model, but hopeful about the potential.

“It’s conceivable that an AI-based model can learn the behavior in these areas more effectively,” he said. “If we get it right, these systems could become an essential part of how we plan for and respond to future fires.”

So far, most public agencies are still in the early stages of adopting these systems. Some counties and utilities have begun installing high-definition camera networks to detect wildfires, but the scale of deployment needed to match the growth of fire seasons remains far off. That has created a market opportunity for startups like Pano AI, which now partners with both public agencies and private companies, including insurance and energy firms.

In wildfire-prone areas, a few minutes of early warning can save lives, homes and millions of dollars, Kastner said. She believes AI can help fix that gap. “Today, AI is helping detect ignitions in real time, reduce false positives and bring together multiple sources of information into a single view that supports faster decision-making,” she said.

In the future, she predicts that AI will expand its role beyond detection, into coordination and resource allocation. This might include providing fire commanders with live overviews of terrain conditions, fuel loads or the likely direction of spread based on changing winds.

At the same time, researchers like Oberai are working to push the frontier further, experimenting with hybrid models that blend physics with AI-trained pattern recognition. These systems might eventually simulate fire dynamics at the block level in suburban zones or predict structure vulnerability with greater precision.

“We’re just beginning to understand what these tools can do,” Oberai said.

There is also a growing effort to apply these technologies outside traditional high-risk zones. Pano AI is now expanding to regions that have historically seen few major fires, but are becoming more vulnerable due to changing precipitation and heat patterns. Kastner says this kind of forward-looking strategy is necessary to adapt to a new climate reality.

“We’re expanding into areas that didn’t use to burn, but are starting to now,” she said. “The threat is growing, and communities need better ways to prepare.”

That preparation will require funding. Pano AI has raised venture capital to grow its footprint, though the company has not disclosed specific numbers. Its pitch is simple: reduce risk through early detection and rapid intelligence. Still, the company’s leaders stress that AI is no silver bullet. It is only as good as the data it receives, the infrastructure it runs on, and the people it supports.

“AI is not replacing humans,” Kastner said. “It’s helping them move faster, with more clarity.”

The future of firefighting may resemble a control room more than a firehouse. It will involve maps, monitors and models, alongside the trucks and hoses. And if researchers like Oberai and entrepreneurs like Kastner are right, that future is already here.

“The fires are changing,” Oberai said. “We have to change with them.”