AI Is Closing the Data Gaps That Have Limited Disaster Prediction for Decades

Fire, flooding, and severe storms all broke records in 2025. But the models that utilities, insurers, and governments rely on to predict disasters can’t keep up.

Some climate scientists call this the “end of stationarity”: the point at which historical baselines for rainfall, water flow, temperature, and extreme weather are no longer relevant. This creates a problem: At a moment when disasters are becoming more frequent, the historical models for measuring and predicting them are becoming less reliable.

AI’s ability to rapidly extract patterns from enormous datasets makes it well-suited for this problem. Utilities, governments, emergency managers, and researchers are using machine learning and large language models to predict risk and respond to disasters with a level of granularity and accuracy that was previously impossible. 

The examples below span wildfires, flooding, and severe storms, but they share a common thread: In each case, AI is filling gaps or improving the accuracy of data that helps those in power stay one step ahead of the next disaster.

AI is Reducing Fire Risk with “Vegetation Intelligence”

Utilities spend about $8 billion per year on vegetation management. Downed lines and contact with vegetation are responsible for several wildfires that have hit the West in recent years, so keeping branches and foliage clear of power lines is one of the most effective ways to prevent ignitions. But when a single utility is responsible for thousands of miles of line across varied terrain, managing that risk—at the right time, in the right priority—is difficult. For context, the Pacific Gas and Electric Company (PG&E), which operates in California, has over 100,000 miles of power lines—enough to wrap around the Earth’s equator over four times.

A growing number of companies offer AI-powered “vegetation intelligence” for this purpose. Overstory, one of the more established, builds tools that tell utility companies where vegetation risk meets ignition risk: the places where dead foliage, downed trees, or fallen branches are most likely to contribute to a wildfire.​

Here’s how it works: Overstory applies its AI tools to satellite and aerial imagery and backs up its data with a team of arborists inputting on–the-ground observations. By combining that information with wildfire risk maps, Overstory can show where trees are too close to circuitry, where trees are at risk of lightning strikes, and where a dying tree could fall onto a power line (known as a “fall-in risk”). The utility can then make informed decisions about where to prioritize trimming and removing trees, and which low-risk areas they can skip.

Overstory says six of the ten largest North American utility firms are among its clients, including BC Hydro, AEP Texas, and PG&E, which declared bankruptcy in 2019 and has invested heavily in wildfire resiliency since.

Inland Power and Light (IPL), a utility based in Washington State, reported reducing hazard trees by 88%. “You can’t manage 5,000 miles at once,” Patrick Larsen, Vegetation and Utility Inspection Manager at IPL, said in a white paper. “But with technology, you can identify what work we need to do now to make the most meaningful difference.”

​Independent data on how much AI-driven vegetation management reduces wildfire ignitions across the industry is still limited, but the utility sector's investment suggests the tools are delivering results worth paying for, with substantial downstream benefits for homeowners, insurers, and local governments.

Hawaii Is Using AI To Warn Communities Faster about Upcoming Disasters

Across the Pacific, Hawaii faces a range of natural disasters, including wildfires, hurricanes, volcanic eruptions, landslides, earthquakes, and flooding. Early detection systems can buy evacuees minutes, and those minutes matter as island communities often have limited evacuation routes.

For decades, the Pacific Disaster Center on Maui has pioneered early warning technology. Recently, it has incorporated AI. The tool’s applications go far beyond Hawaii: The center's DisasterAWARE platform aggregates data from thousands of sources, like satellite feeds, weather models, live cameras, and infrastructure data, into a dashboard used by more than 20,000 disaster management and humanitarian practitioners worldwide. The platform can estimate impacts to population and critical infrastructure within minutes, including a demographic breakdown of who’s exposed and what the likely humanitarian needs will be.

For wildfire specifically, the PDC receives remotely sensed satellite data through a NASA partnership. It integrates that data into DisasterAWARE to track where fires are burning and how they're spreading. Then it enriches that information with AI to generate alerts and analytics in near real time.

“Saving lives and protecting communities from wildfires depends on decisions made in moments of uncertainty,” said Dr. Erin Hughey, Deputy Executive Director of PDC, in a statement. “That's why it's essential for emergency managers to have trusted data, actionable insights, and rapid analysis of risks all in one place.”

The island of O'ahu had been moving toward exactly that kind of system: A new $1 million federal grant was set to fund an AI-driven hazard prediction and warning system. The system ingests data from the National Weather Service, local stream monitors, cameras, and real-time reports from utilities and city crews, along with “digital twins”: virtual replicas of specific infrastructure sites, built from drone footage and real-world data, that let emergency managers simulate disaster scenarios before they happen.

"We can go to critical infrastructure around O'ahu, and then I can use a digital twin to ask, 'What happens if there's a landslide?'" Randal Collins, director of the Honolulu Department of Emergency Management, told Honolulu Magazine. "We can then better coordinate a plan."​

However, Collins announced in April that the federal funding had been frozen, putting the project on hold—a reminder that early warning infrastructure built on federal grant cycles can be unreliable.

Google Is Filling A Decades-Old Gap in Flood Data

One in eight Americans is exposed to flood risk, and that number is only rising. Strong flood prediction tools can both alert residents to flash flood risk and help cities, businesses, and investors make informed resilience decisions, but the patchwork of flood data has made accurate prediction challenging. While some natural disasters, like earthquakes and tsunamis, are tracked by global sensor networks, floods don’t have a standardized method or repository for data collection and tracking. This lack of historical data for flash floods has created a “data desert” that limits accurate forecasting.

In March 2026, Google announced a new project called Groundsource that uses Gemini AI to mine the world's news for flood data. The system analyzes news reports where flooding is a primary subject, translates text from 80 languages, and then uses Gemini to classify whether an article describes an actual, ongoing, or past flood. It then maps the impacted neighborhoods and streets using Google Maps.

The result is a dataset of 2.6 million historical flood events spanning more than 150 countries. Using the Groundsource dataset, Google trained a new AI model that has made progress toward predicting flash floods in urban areas up to 24 hours in advance. Those forecasts are now freely available through Google's Flood Hub.

That free availability is itself an interesting market signal. Insurers and catastrophe modelers—the firms that quantify disaster risk for the insurance and finance industries—currently pay for proprietary disaster risk data. Free, high-quality alternatives could reshape who has access to that information, giving real estate investors, businesses, and local governments better tools to price exposure and make land-use decisions in flood-prone areas.

Google says that this methodology—turning unstructured data from articles, reports, and bulletins into a common format for modeling—could be extended to other hazards that have historically lacked precise records, like droughts, landslides, and avalanches.

Better Data is Stretching Severe Weather and Hail Forecasts From Hours to Weeks

Fires and floods are far from the only disasters getting the AI treatment. A new AI tool built by the National Science Foundation can now help forecasters identify the potential for severe weather outbreaks a week in advance, with plans to expand that to two weeks. And hail, long one of the hardest hazards to model because damage accumulates across many smaller events, is also drawing AI business formation and investment: A startup called FLASH Weather AI recently launched a hail prediction model capable of forecasting hail size and arrival time up to 55 minutes ahead.

These efforts are all connected by a shared premise: that the data gaps limiting how well utilities, governments, and emergency managers make decisions about risk are closeable—and that closing those gaps reduces risk and boosts preparedness. As disasters push into new geographies and old models lose their predictive power, resilient decision-making requires tools that can keep pace with the change.

Have thoughts to share on this piece, or want to add your voice to the conversation? Reach out!