Welcome to Ground Truth, OmniGeo’s publication on the intersection of geospatial AI and physical risk — for the people building in Earth observation and the underwriters, asset owners, and risk teams who have to trust them.

What a year of listening taught us

In late 2025, Headwaters Economics, the wildfire analytics firm Pyrologix, and the U.S. Fire Administration set out to take stock of how the industry actually models wildfire risk. They started with more than 150 risk models, narrowed the field to the 59 that met their criteria, and then interviewed 30 of the people who do this work for a living: fire physicists, structural engineers, catastrophe modelers, and insurers.

The experts disagreed about almost everything, from whether physics-based or simplified operational models are the future to how close anyone is to modeling the way fire jumps from one structure to the next. On one point, they converged almost unanimously. As one interviewee put it, “models are only as good as the data inputs.” A field full of people who build models for a living, concluded that the single highest-leverage gap in wildfire risk is not another model. It is a scalable, validated measurement of the physical condition of individual buildings and the land immediately around them.

We did not need the report to tell us that, though it was striking to watch a field arrive at the conclusion we had reached the hard way. One of us lost a home in the Marshall Fire, the suburban firestorm outside Boulder that destroyed more than a thousand structures in an afternoon. Over the past year, we have sat across the table from carriers, MGAs, reinsurance brokers, and catastrophe modelers, and a few themes surfaced in nearly every conversation. Together, they became the brief for the product we are launching today.

The first is that hazard is not the same as vulnerability. The industry has spent two decades getting very good at modeling fuel loads, wind fields, drought, and fire spread, none of which explains why one house on a street burns to the foundation while the one next door is untouched. That difference is vulnerability, and it lives at the parcel level: the roof material, the vegetation within five feet of the wall, the canopy over the eaves, and whether the homeowner actually cleared what they were told to clear. These are measurement questions, not modeling questions, and they have historically been expensive, manual, and nearly impossible to keep up to date.

The second theme was sharper, and it came up most forcefully from the people closest to the underwriting desk. The capability nobody can buy today is dynamic, near-real-time fuel and vegetation moisture, paired with the ability to verify mitigation after the fact. A carrier tells a homeowner to clear the vegetation in Zone 0, the five feet closest to the structure. The homeowner does the work. And then nothing happens, because there is no way to confirm it until the next aerial flyover, perhaps a year away. In a market increasingly required to give credit for mitigation, that blind spot is expensive.

The third theme was about delivery. Almost no one wanted another black-box score. The largest carriers want raw, explainable features for their own pricing models, smaller carriers want a simple number they can act on, and the data providers in between want differentiated inputs to upgrade what they already sell. The lesson was to serve as the measurement layer that feeds everyone else’s models, rather than as another score. And that, every serious buyer agreed, is proven in only one way: correlation with claims, validated against post-fire damage records, such as California’s DINS database, that show which homes actually burned and which survived.

Underneath all of it, regulation is forcing the issue faster than the data can keep up. California’s defensible-space rules under AB-3074 and Colorado’s HB25-182 increasingly require insurers to evaluate the vegetation immediately around a structure, even as regulators warn that carriers cannot act on or non-renew based on an image showing nothing more than cosmetic wear. The market is now legally obligated to understand parcel-level conditions and mitigation, using data that largely does not yet exist at scale.

That gap is the reason we built OmniGeo. And it is the reason we are launching our first product today.

Introducing OmniFire

OmniFire is a property-level wildfire risk API for underwriters, MGAs, reinsurers, and the risk platforms that serve them. It is spatial AI for physical risk: it turns frequently refreshed, high-resolution satellite imagery into explainable, parcel-level features that describe how an individual structure is likely to perform when fire arrives. What is the roof made of? How much vegetation surrounds the building? How dry is that vegetation? Is there a canopy hanging over the roof? Does the defensible space around it meet current regulations?

It is built to fill the exact gap the industry described to us. Rather than another regional hazard map, another infrequent aerial flyover, or another costly drone inspection that cannot scale to millions of policies, OmniFire delivers measurements: structured, auditable, and traceable all the way back to the image they came from. Our initial coverage focuses on the California and Colorado counties with the highest wildfire exposure and the densest concentrations of insurable property, with the rest of the western United States to follow.

From Hazard Scores to Material Intelligence

Consider two neighbors. A conventional wildfire hazard map assigns both homes the same regional score because they sit on the same hillside, in the same wind corridor, and are subject to the same drought. But one has a metal roof, a clean perimeter, and well-watered landscaping, while the other has a wood-shake roof, dense brush against the siding, and dry fuel under an overhanging canopy. Their regional hazard is identical. Their vulnerability could not be more different. The hazard map cannot tell them apart, and that distinction, multiplied across a portfolio, is the difference between pricing risk and guessing at it.

OmniFire is built around a single principle that came straight out of those discovery conversations: provide features, not scores. We do not compress a property into one number and ask you to trust it. We deliver the underlying parcel-level observations, roof material, defensible-space vegetation, tree proximity, canopy overhang, vegetation moisture, and more, each one carrying full provenance back to its source image, acquisition date, and processing pipeline. You retain complete control over your underwriting models, pricing, and regulatory filings. We supply the material intelligence that powers them.

What makes that possible is a fundamentally different way of reading imagery. Most computer vision is trained to answer “what is this?” OmniFire is built to answer “what is this made of?”, recovering the physical signal of material, moisture, and condition that ordinary RGB imagery leaves on the table. We will go deep into the science with our co-founder, Dr. Ophir Almog, in a future post. For now, here is what it produces.

What's in the First Release

The first version of OmniFire delivers four core categories of wildfire intelligence, each one a measurement rather than an inference dressed up as a score.

Roof material is one of the strongest single predictors of whether a structure survives a wildfire, and OmniFire classifies the dominant roof material for every building, separating asphalt and concrete, metal, tile, wood, and built-up tar-and-gravel surfaces. This is a signal we validate at roughly 90 percent accuracy against county records, and because it is spectral rather than visual, it correctly identifies two metal roofs of different colors as the same material, whereas an RGB system might call them different.

Defensible-space vegetation answers the question that the new regulations are built around. OmniFire measures vegetation density in the standard defensible-space zones around every structure, reporting separately on Zones 0, 1, and 2, as well as the canopy overhanging the roof. Instead of a subjective field note, an underwriter gets a quantitative, repeatable number for each zone.

Tree density and tree-to-structure distance isolate the fuel that matters most. Trees carry more mass, more canopy, and more ember potential than low vegetation, so OmniFire reports tree canopy coverage by zone, canopy directly over the roof, and the distance from the nearest tree to the building, letting a carrier see not just how much vegetation surrounds a home but how close the dangerous part of it sits.

Vegetation moisture captures flammability over time. OmniFire computes a Normalized Difference Moisture Index from Sentinel-2 imagery to track seasonal water stress and adds a Live Fuel Moisture Content proxy, built from optical, radar, climate, and topographic data, to estimate fuel moisture dynamics throughout the year. Together, they describe ignition conditions, the dynamic layer that the market told us it has been missing.

Every one of these features is delivered with full provenance, traceable to the image, the date, and the pipeline that produced it. And in the validation that this industry trusts most, our signals showed roughly 97 percent agreement with California’s post-fire damage inspection records on which structures survived and which were destroyed, outperforming the standard public datasets.

What's Coming Next

OmniFire is built to expand into a broader spatial AI platform for physical risk, with several capabilities already in development. Near-term additions include tree-specific detection, an expanded roof-material taxonomy, vegetation-change monitoring to verify mitigation over time, roof condition and estimated roof age, deck detection and deck material, ember-resistant hardening features, ladder-fuel identification, and surface-type analysis within the defensible-space zones. Several of these, particularly change monitoring and ladder fuels, came directly out of the discovery conversations, and they point toward the same destination: a continuously updated, parcel-level picture of structural vulnerability and the mitigation that changes it.

Work With Us

If you underwrite or model wildfire risk in California, Colorado, or the broader western United States, we would like to put OmniFire in front of you with a retrospective property test: give us a set of addresses, and we will return parcel-level signals you can compare against what you already know. Let us show you what OmniFire sees at www.omnigeo.ai/omnifire.

The measurement layer behind Ground Truth

OmniGeo’s first product, OmniFire, closes the earth measurement gap for insurers in wildfire-prone areas. OmniFire turns standard imagery into parcel-level measurements of what a property is actually made of, including roof material, defensible space, and fuel moisture. It’s vulnerability measured at scale and it’s live today. If you underwrite, price, or manage wildfire risk, visit www.omnigeo.ai/omnifire to learn more.

In Altadena this past January, the Eaton Fire left behind the image that has come to define the modern wildfire: one neighborhood reduced to ash, and right in the middle of it, seemingly untouched, a home that survived.

Now, it turns out, the difference between destruction and survival comes down to more than just luck. And we now have the data to prove it.

Over the past year, a remarkable run of research has converged on the same conclusion: the physical condition of an individual property, what the roof is made of, what’s growing in the first five feet, and whether the vents are screened, are the strongest determinants of whether that home survives. Home hardening and mitigation work, and the science behind it is now hard to argue with.

This raises a new question: if survival comes down to parcel-level physical detail, can we actually measure that detail at scale? Because most of the tools the industry relies on today cannot.

The 52% finding

The most rigorous evidence yet comes from a study led by UC Berkeley, published in Nature Communications in August 2025. Maryam Zamanialaei, Michael Gollner, and colleagues combined CAL FIRE’s ground-truth damage inspections from five of California’s most destructive fires (the 2017 Tubbs and Thomas, 2018 Camp, 2019 Kincade, and 2020 Glass fires) with physics-based fire and ember simulations across roughly 47,000 structures. A machine-learning model trained on the result predicted structure survival with 82% accuracy.

The headline finding is that hardening and defensible space, applied together, can dramatically reduce loss. In the study’s most aggressive scenario, which combined hardening the structure with clearing defensible space across the first 30 feet, modeled survival nearly doubled to 48%, and the authors summarize the combined effect as a “hypothetical 52% reduction in losses.”

“I view this as really powerful evidence that the mitigation measures that are available to us, hardening and defensible space, actually have some real-world effectiveness,” Gollner said, regarding their study. And on the limits of what we can control: “We can’t always change the spacing between structures or the exposure from flames and embers. But even within those limitations, we still have the power to cut the destruction in half, if not more. That is very powerful.”

The model also ranked what matters. Structure separation distance was the single strongest driver of loss, followed by exterior siding material, then year built, then flame exposure. The team found it was “most impactful to make changes both to the structure itself and surrounding fuels, especially vegetation and other flammable materials within 1.5m (5ft) of the structure.” Clearing that first five-foot zone alone, the contested “Zone 0,” could cut structure losses by 17%.

Note what every one of those variables has in common: siding material, roof construction, vent screens, vegetation within five feet. They are physical, material properties of a specific parcel.

From Survival Rates to Insurability

What the Berkeley team showed in physics, California’s regulators just translated into dollars. In March 2026, the California Department of Insurance and the NAIC released the first study to model how community-wide rebuilding to a hardening standard affects Average Annual Loss, the metric carriers lean on most heavily when deciding whether to write a policy at all.

Using Moody’s wildfire catastrophe model, they found that rebuilding the roughly 30,000 homes inside the Palisades and Eaton burn perimeters to the IBHS Wildfire Prepared Home standard would cut average losses by about a third (31% for the Base standard, 35% for Plus) at a marginal cost of roughly 3% per home. That is the difference, the Department argued, between a community stuck on the FAIR Plan and one a competitive private market will actually serve.

“Every home rebuilt to the Wildfire Prepared Home standard is a home that is safer for the family inside it, safer for its neighbors, and more likely to remain insurable for decades,” said Commissioner Ricardo Lara. Several of California’s largest carriers have now committed to writing policies for homes that earn the designation. Mitigation is no longer just a safety story; it has become an underwriting input.

What the Lab Confirms

The Insurance Institute for Business & Home Safety (IBHS) has spent more than a decade burning this question down to specifics in its research center. Their post-fire forensics and controlled ember tests keep pointing at the same culprits. Embers, not the flame front, are the leading cause of home ignition, and IBHS finds that an ember-resistant buffer in the first five feet can cut a home’s risk of igniting roughly in half.

“Time and time again in our post-fire analysis of the Los Angeles County wildfires, we saw the damage done by combustible items like plastic garbage cans right up against a home,” said Steve Hawks, IBHS Senior Director for Wildfire and a 30-year CAL FIRE veteran. “The bottom line based on the latest research is this: a limited area that is completely noncombustible prevents ember ignition.” Their checklist of what actually moves survival is short and concrete: a Class A fire-rated roof, ember-resistant vents, at least six inches of noncombustible material at the base of exterior walls, and a clear five-foot zone.

The Measurement Problem

So here is where the science runs into a wall. Every credible source above says the same thing: survival is decided by material, parcel-level facts, including roof composition, siding, vent type, the moisture and fuel of vegetation within feet of the wall, and whether mitigation has actually happened. Yet the dominant way the industry “sees” property risk today is aerial and satellite RGB imagery, and RGB has a fundamental blind spot, because it captures color and shape rather than material composition. A camera can make a best guess that a roof looks like tile, but it cannot tell you whether the roof is truly combustible, measure the moisture in the vegetation against the wall, or determine whether a vent is ember-resistant. It registers objects, while the physics that actually decides survival stays invisible to it.

The usual workaround is to train computer-vision models to label those objects, an expensive and brittle path. Building and labeling them costs millions, and as we wrote last time [insert link], even the most sophisticated foundation models degrade the moment the world shifts underneath them, whether that is a new region, a different sensor, a different season, or a post-event scene. Worst of all for this use case, aerial imagery is a snapshot, often a stale one, and it can’t price the thing the science says matters most: whether a property is more or less vulnerable after mitigation than it was before.

That is the gap. The science has settled whether parcel-level hardening works; what it has not solved is how to measure the physical condition of millions of homes accurately, by material, and over time. We started OmniGeo because that measurement problem, not another model, is the real bottleneck. Mitigation that no one can see can’t be priced, credited, or rewarded, and a home that can’t prove it’s been hardened won’t get the benefit of the doubt. And systematically rewarding mitigation is the key to a more resilient future and insurance ecosystem.

The measurement layer behind Ground Truth

OmniGeo’s first product, OmniFire, closes the earth measurement gap for insurers in wildfire-prone areas. OmniFire turns standard imagery into parcel-level measurements of what a property is actually made of, including roof material, defensible space, and fuel moisture. It’s vulnerability measured at scale and it’s live today. If you underwrite, price, or manage wildfire risk, visit www.omnigeo.ai/omnifire to learn more.

Welcome to Ground Truth, OmniGeo’s publication on the intersection of geospatial AI and physical risk — for the people building in Earth observation and the underwriters, asset owners, and risk teams who have to trust them.

There's a race underway right now to do for the Earth what Google once did for the web: make it searchable. You type a question and get back an answer grounded in satellite imagery. The building blocks are everywhere, from Google's planet-scale embeddings to IBM and NASA's Prithvi, NVIDIA's Earth-scale models, and open foundation models like Clay and TerraMind multiplying by the month. The pitch is seductive: compress the planet into vectors, and anyone, an analyst, an app, an agent, can query the world in seconds.

I truly believe in where this is heading. But there's a problem the demos tend to skip over, and it isn't a small one. For all the talk of seeing the planet, AI doesn't actually see the Earth at all; it pattern-matches pixels, and we now have hard evidence of just how far that gap goes.

What the New EarthShift Benchmark Found

Last month, researchers at Arizona State released EarthShift, the first serious public benchmark built to measure not how well geospatial models perform, but how well they hold up when the world changes underneath them. They tested 8 leading foundation models, among them Clay, Prithvi, TerraMind, and DINOv3, across 11 tasks and five kinds of real-world "shift." The results point to some critical issues with these existing approaches.

Five ways the world shifts, and which ones break the models

Lead author Kelsey Doerksen summed it up in a single word: GFMs are “brittle.” Here are the five shifts that successfully unmoor a deployed model:

Sensor shift: Train a model on imagery from one satellite, then run it on another, say Sentinel-2 to Sentinel-1, or optical to radar. Both see the same ground through very different instruments, with different spectral bands and radiometry. The degradation from this type of shift was the most severe in the benchmark, with accuracy collapsing by up to 60%.

Geographic shift: Train in one part of the world and deploy somewhere else entirely. A field-mapping model trained on Germany’s large, orderly parcels falls apart over the smallholder plots of Cambodia, where field sizes, crops, and landscape look nothing alike. As you’d expect, the bigger the leap, the worse it gets.

Scale shift: This one is about resolution: the same scene at very different levels of detail. Picture a model trained on coarse public imagery, where each pixel covers ten meters of ground, then pointed at sharp commercial imagery measured in tens of centimeters. A rooftop that was a single fuzzy smudge becomes a richly textured surface, and the model’s grasp of object size and texture breaks down. Even systems with strong in-distribution scores lost a third to two-thirds of their performance.

Data source shift: This is the subtle one, and the most unsettling. Everything appears to match: the same task, sensor, resolution, and label classes. Only the provider or processing pipeline behind the images has quietly changed. Imagine two vendors both selling “analysis-ready” Sentinel-2 data, where one applies a different atmospheric correction, so pixels a human would swear are identical carry different values. The model still drops, which is what makes this case so dangerous: nothing looks wrong, and yet the answer is quietly off.

Temporal shift: Finally, the same place observed at a different moment in time. Think of a cornfield in June, lush and green, and then again in October after the harvest, reduced to bare soil and stubble. Same land, completely different pixels. This was the one shift the models mostly handled, which is itself the tell: they cope with the gentlest change and stumble on the four that define real deployment.

On average, these models shed around 20% of their performance the moment they were pushed beyond their training set. The real takeaway is more uncomfortable: models built specifically for Earth were no more robust than generic image models trained on ImageNet, and in several cases no better than a randomly initialized network. Domain-specific pretraining, the premise behind “large Earth models,” bought essentially no robustness, and fine-tuning didn’t fix it; sometimes it made things worse. As one GeoAI engineer put it, “models don’t fail loudly. they drift.”

And it isn’t only the models that see the Earth; it is also the ones you talk to. The whole promise of a searchable planet rests on vision-language models, and a second benchmark released in March, OmniEarth, ran those through 28 geospatial tasks and found they “still struggle with geospatially complex tasks," sometimes answering from language patterns rather than the image itself. A confident answer not grounded in the pixels is the textbook definition of geospatial hallucination.

Why the People with Real Money On the Line Aren't Running on Embeddings

When a decision is truly consequential, you don't see raw geo-embeddings making the call. There is always a human in the loop re-validating the output before anyone acts. That isn't mere conservatism, because the EarthShift results suggest it is the rational response. Much of what gets marketed as "Earth AI" is, underneath, a wrapper around models that don't generalize past their training distribution, and the industries with the most to lose feel that instinctively. As Melissa Rosa put it in a sharp essay on scaling Earth observation, a model can be accurate on average and still end up "confidently wrong in the case that matters most." Embeddings can tell you two places look similar, but similarity is not the same thing as attribution. Decision-makers need to know what actually changed, why it matters, how reliable the signal is, and what to do next, with calibrated confidence, much the way a 70% chance of rain is a number tested against millions of real outcomes.

What actually closes the gap

The path forward is not a bigger model. EarthShift found that scale doesn’t buy robustness: model size had no correlation with how a system held up out of distribution. The gap closes instead with three things the current wave mostly skips:

Measurement built for stability and comparability: utilizing a signal that means the same thing across different sensors, seasons, and geographies, rather than a representation that drifts the moment conditions change.

Attribution grounded in physical reality, not statistical similarity: answers tied to what is physically there and why it matters, validated against real ground truth, rather than a claim that “this vector is near that vector.”

Calibrated confidence: a system that tells you not only what it sees, but how much you should trust it under these specific conditions, improved through feedback loops between prediction and observed outcome.

These may seem like overly theoretical fixes, but over the coming weeks, we will dive into some of the technological innovations pioneered by our team at OmniGeo and how we believe we can meet these criteria in completely novel ways.

All of this is harder than building another foundation model; it demands fluency in both the underlying physics and the operational reality of the decision. But it is the only version an underwriter, an emergency manager, or a defense analyst can stand behind, and the only one that doesn’t end with an insurance customer losing their coverage based on a geospatial hallucination.

The web, after all, was made of links, and Google won by indexing them. The Earth is not made of links. It is made of atoms, and you cannot compress your way past that. The companies that define the next decade of GeoAI won't be the ones that make the Earth searchable. They'll be the ones that can hand you an answer you'd actually bet on when the stakes are real.

The measurement layer behind Ground Truth

OmniGeo’s first product, OmniFire, closes the earth measurement gap for insurers in wildfire-prone areas. OmniFire turns standard imagery into parcel-level measurements of what a property is actually made of, including roof material, defensible space, and fuel moisture. It’s vulnerability measured at scale and it’s live today. If you underwrite, price, or manage wildfire risk, visit www.omnigeo.ai/omnifire to learn more.

Want to learn more about the future of Earth AI? Visit us at OmniGeo.

On December 30, 2021, the Marshall Fire destroyed more than 1,000 homes in Boulder County in a matter of hours. Not days. Hours. For many Coloradans, the Marshall Fire shattered a deeply held assumption about wildfire risk: that wildfire is primarily a forest problem. It wasn’t.

The Marshall Fire was largely a grass-fueled, wind-driven urban conflagration. Once structures began igniting, homes themselves became fuel. Embers traveled miles. Entire neighborhoods burned despite being far from dense forests. Yet nearly five years later, many of the public wildfire risk maps used by homeowners, policymakers, insurers, and planners still struggle to represent this reality.

A recent criticism of the Colorado State Forest Service’s Wildfire Risk Viewer highlighted exactly this concern. The map correctly identifies wildland fire intensity across the landscape but appears less capable of representing how risk evolves once fire enters the built environment. The criticism may be blunt. The underlying issue is not. It’s one of the most important challenges in wildfire science today.

The Problem Isn't Forest Fire Modeling

To be fair, the Colorado Wildfire Risk Viewer was never designed to be a parcel-level urban fire model. It is a valuable public resource that helps homeowners understand: fire intensity potential, burn probability, historical fire occurrence, broad landscape-level hazards.

Those are important inputs. But they are not sufficient for understanding what happened in Louisville, Superior, Lahaina, Paradise, or increasingly many communities across the American West. Because once wildfire reaches neighborhoods, the risk equation changes. The fuels are no longer just: trees, shrubs, grasslands. The fuels become: roofs, fences, decks, sheds, vehicles, landscaping, building-to-building spacing, construction materials. In other words:

The built environment becomes the fuel model. And most wildfire risk maps still have limited visibility into it.

The Missing Layer: What Things Are Made Of

Modern wildfire maps are incredibly sophisticated at characterizing vegetation.
We have:

• Fuel models

• Canopy height models

• Vegetation density estimates

• Drought indicators

• Weather forecasts

• Fire spread simulations

What we often lack is an equally detailed understanding of the structures those fires threaten. For example: two homes may look identical from a traditional risk model.
But one may have:

• Class A asphalt shingles

• Stucco siding

• Metal fencing

• Five feet of defensible space

While the other has:

• Wood shake roofing

• Combustible fencing

• Dense ornamental vegetation

• Attached wooden structures

From a wildfire perspective, these are not the same risk. Yet many existing systems cannot distinguish them at scale.

Wildfire Risk Has Become a Materials Science Problem

Historically, wildfire risk assessment focused on: “How likely is fire to arrive?” Today, insurers, utilities, and communities increasingly need to answer: “What happens after it arrives?” That question depends heavily on materials. Not just geometry. Not just location. Not just vegetation. Materials. The difference between: asphalt and wood shake roofs, green vegetation and cured vegetation, metal fencing and cedar fencing, moist fuels and dry fuels often determines whether a structure survives or is lost. The challenge is that most geospatial systems still primarily understand the world through shape and appearance. Wildfire behavior increasingly requires understanding the world through composition.

We Need Fuel Maps for Communities, Not Just Forests

Imagine if every parcel had an evolving digital fuel profile. Not simply: high risk, medium risk, low risk but: roof material, roof condition, vegetation type, vegetation moisture, defensible space quality, fuel continuity, building spacing, changes through time. Updated continuously from Earth observation systems. Now imagine combining those data with: weather forecasts, topography, fire behavior models, historical losses. That begins to look less like a traditional wildfire map and more like a living model of community resilience.

The Technology Is Finally Catching Up

For years, this vision was impossible. The imagery wasn’t available. The computing power wasn’t available. The AI wasn’t available. Today, all three are converging. A new generation of Earth observation systems, geospatial AI, spectral analysis techniques, and large Earth models are beginning to move beyond identifying objects toward identifying materials and conditions.

Instead of merely recognizing a roof. They can begin asking: What is the roof made of? How has it changed? What condition is it in? What surrounds it? How combustible are those surroundings? That shift, from seeing shapes to understanding materials, may prove as important for wildfire risk as weather forecasting was for hurricane prediction. The broader Earth observation community is increasingly moving in this direction, with new foundation-model approaches focused on generating richer representations of the physical world from satellite imagery and multi-modal geospatial data.

The Next Marshall Fire Is Already Being Modeled Just Not Completely.

The uncomfortable reality is that many current wildfire maps are still optimized for the fires of the past. The fires we’re seeing now increasingly involve: extreme winds, human-built fuel networks, structure-to-structure ignition, rapid urban fire spread.

These events sit at the intersection of ecology, meteorology, engineering, and materials science. No single map layer can capture them. No single model can capture them. And no single agency should be expected to solve the problem alone. But if we learned anything from Marshall Fire, it should be this: Wildfire risk doesn’t end where the forest ends.

That’s where a different kind of risk begins. The future of wildfire intelligence won’t come from better vegetation maps alone. It will come from understanding the complete fuel landscape, from forests to fences, from shrubs to shingles, from ecosystems to neighborhoods. Only then will we start building risk models that reflect how modern wildfires actually behave. And only then will we stop being surprised by disasters that, in hindsight, were hiding in plain sight.

The measurement layer behind Ground Truth

OmniGeo’s first product, OmniFire, closes the earth measurement gap for insurers in wildfire-prone areas. OmniFire turns standard imagery into parcel-level measurements of what a property is actually made of, including roof material, defensible space, and fuel moisture. It’s vulnerability measured at scale and it’s live today. If you underwrite, price, or manage wildfire risk, visit www.omnigeo.ai/omnifire to learn more.

Welcome to Ground Truth, OmniGeo’s publication on the intersection of geospatial AI and physical risk — for the people building in Earth observation and the underwriters, asset owners, and risk teams who have to trust them.

In late 2025, Headwaters Economics, the wildfire analytics firm Pyrologix, and the U.S. Fire Administration published a two-part study with a deliberately boring title: Wildfire Risk Indices and the Built Environment. Part one is an inventory: the authors started with more than 150 wildfire risk models, narrowed down to 59 that met their criteria, and scored each on how well it captured the built environment and whether it could run at national scale. Part two is the interesting part: they interviewed 30 subject matter experts, including fire physicists, structural engineers, catastrophe modelers, and insurers who do this work, and asked them how models are broken today.

Buried in the findings is the reason we started OmniGeo.

This field of experts largely agrees that one of the most important datasets needed to reduce wildfire losses does not yet exist. That matters because wildfire is no longer just a forestry problem. It’s an insurance problem, a mortgage problem, a municipal finance problem, and increasingly, a household balance sheet problem.

The question isn’t whether a wildfire will burn in a particular area. We have a good understanding of this. Any area that abuts a moderate or higher wildfire risk area has the potential to burn, especially in areas with high wind potential. The question now is which homes burn, which survive, and whether we can predict the difference before the fire arrives. This study is remarkably transparent in asserting that our wildfire models still struggle to do that.

The Wildfire Modeling Paradox

Over the last two decades, wildfire modeling has become very sophisticated. We can model fuel loads, topography, wind fields, drought stress, ember transport, and fire spread across landscapes at scales that would have seemed impossible a generation ago. Yet some of the most destructive fires in recent history have exposed a persistent blind spot: the built environment.

As Headwaters Economics notes, most existing wildfire models remain fundamentally optimized for understanding how fire moves through vegetation, not how it propagates through neighborhoods. Structure-to-structure ignition, building materials, parcel conditions, and mitigation actions remain poorly represented in most operational risk models. That distinction may sound academic, but it isn’t.

When a fire enters a neighborhood, the rules change. A burning cedar fence behaves differently from a ponderosa pine. A wood shake roof behaves differently from concrete tile. A home with five feet of defensible space behaves differently from one with vegetation touching the structure.

These differences often determine single losses measured in hundreds of thousands or millions of dollars. Yet they remain surprisingly difficult to measure consistently across entire states.

What the Experts Said

The second phase of the Headwaters study interviewed 30 wildfire risk experts across disciplines. The findings were fascinating. There was disagreement about many things: whether behavioral change or better models matter more, whether physics-based models or simplified operational models are the future, and how close we are to accurately modeling structure-to-structure fire spread. But there was striking agreement around one issue: a lack of data.

As one interviewee put it: “Models are only as good as the data inputs.” That sounds obvious, yet the implications are enormous. Among the report’s most important recommendations was the need for better datasets that capture:

• Building-level characteristics

• Parcel-level conditions

• Ember transport dynamics

• The impact of mitigation activities on fire behavior

The report explicitly identifies remote sensing as one of the most promising paths for developing these datasets at scale. One of the most comprehensive recent reviews of wildfire risk modeling concluded that one of the highest-leverage opportunities isn’t another fire model, but rather better measurement.

Hazard Is Not the Same as Vulnerability

This distinction gets lost surprisingly often. Wildfire risk is typically described as a combination of hazard, exposure, and vulnerability. Combine the three, and you can measure the risk and size of loss.

Hazard is where most attention goes. How likely is a fire? How intense could it become? How fast might it spread? These questions are critically important, but they don’t explain why one home survives while another burns. That is defined by vulnerability.

Vulnerability is measured at the parcel level. What is the roof made of? How much vegetation exists within five feet of the structure? Has mitigation occurred? Has the roof deteriorated? Did the homeowner replace wood fencing with non-combustible materials? Did defensible space improve since the last inspection? These are fundamentally measurement questions, not modeling questions.

However, the challenge is that they have historically been expensive, manual, and difficult to update.

The Cost of Not Knowing

The repercussions of this can be seen across the risk value chain. Insurance carriers raise premiums. Others stop writing policies entirely. Homeowners invest thousands of dollars in mitigation without knowing whether it materially changes risk or the cost of their insurance policy. Communities struggle to prioritize interventions and invest in neighborhood-level mitigation. Regulators push for more granular understanding while insurers often rely on incomplete information.

The result is a market operating with partial visibility. And partial visibility is expensive. One of the most interesting observations from the Headwaters report is that current wildfire models often face a tradeoff. Models that scale nationally often lack the precision of the built environment. Models with detailed precision often struggle to scale nationally.

Solving this bottleneck requires capturing that precision at a national scale. The problem is not a lack of imagery, compute, or machine learning. It is a lack of scalable, validated measurements of the physical conditions that actually influence loss.

A Remote Sensing Opportunity

The report repeatedly points toward advances in remote sensing, machine learning, and improved data collection as key enablers of the next generation of wildfire risk models. That observation mirrors something we’ve seen repeatedly across insurance, infrastructure, utilities, and climate risk.

The world has become incredibly good at collecting images. We’re still surprisingly bad at extracting physical intelligence from them. The challenge is no longer obtaining pictures of a property, but rather understanding what those pixels reveal about materials, condition, vegetation, and vulnerability. The opportunity is to convert imagery into measurement using novel methods. And that’s what we’re doing at OmniGeo.

The Measurement Decade

The first generation of wildfire risk models focused on understanding fire. The next generation will need to understand properties: not just where they are, but what they’re made of, how they’re changing, and whether mitigation actually works.

The Headwaters report does not prescribe a specific technical solution, but it does clearly identify where our knowledge breaks down. The industry needs better data on buildings and parcels, better visibility into mitigation, and scalable ways to collect that information.

That may turn out to be one of the most important opportunities in wildfire risk over the next decade. Because before we can accurately model vulnerability, we first have to measure it.

And right now, that’s still the hardest part.

The measurement layer behind Ground Truth

OmniGeo’s first product, OmniFire, closes the earth measurement gap for insurers in wildfire-prone areas. OmniFire turns standard imagery into parcel-level measurements of what a property is actually made of, including roof material, defensible space, and fuel moisture. It’s vulnerability measured at scale and it’s live today. If you underwrite, price, or manage wildfire risk, visit www.omnigeo.ai/omnifire to learn more.

Welcome to Ground Truth, OmniGeo’s publication on the intersection of geospatial AI and physical risk, for the people building in Earth observation and the underwriters, asset owners, and risk teams who have to trust them.

On December 30, 2021, a grass fire ignited near the Marshall Mesa trailhead between Boulder and Superior, Colorado. Within hours, driven by hurricane-force winds and exceptionally dry conditions, it became the most destructive wildfire in Colorado history. More than 1,000 homes were destroyed. Entire neighborhoods disappeared in a single afternoon. One of those homes belonged to Scott Brave, my long-time colleague and friend, and our Co-Founder & Chief AI Officer.

The fire ultimately stopped roughly a mile and a half from our house. Close enough that we packed our car and left without knowing whether we would have a home to return to.

The Marshall Fire wasn’t Paradise, California. It wasn’t a remote mountain town surrounded by dense forest. It tore through suburban neighborhoods, business parks, schools, and shopping centers. Yet more than a thousand homes burned.

The event quickly became a case study in what fire scientists call urban conflagration: an increasingly important phenomenon where extreme weather, drought, vegetation, building materials, and structure-to-structure ignition combine into a cascading disaster. Areas not perceived as high-risk wildfire zones suddenly became the center of one of the country’s costliest disasters.

We have built some of the most sophisticated models to understand climate risk. We can simulate hurricane tracks, estimate flood probabilities, and project wildfire behavior. As I looked at the charred landscape by my home, I started to wonder: how could this happen? How did we miss this?

A search for answers

This was not a new field of study for me. Through my prior work with NatureX and the Nature Tech Collective, I had spent years studying physical risk, including wildfire, drought, biodiversity loss, insurance markets, and resilience, in conversations with insurers, reinsurers, catastrophe modelers, investors, regulators, and scientists. After enough of these discussions, a pattern had emerged.

While everyone was focused on modeling the future, there was a bigger, more urgent gap that linked back to how an event like the Marshall Fire could occur. This gap was illustrated by the questions I heard come up again and again. An insurer wanted to know whether a property had become more vulnerable to wildfire since its last inspection. A lender wanted to understand the physical exposure of a mortgage portfolio. A utility wanted to know where vegetation was creating operational risk. A conservation organization wanted to understand whether an ecosystem was improving or deteriorating.

The missing information was remarkably similar. Therein lies the problem: we have difficulty measuring the current condition of the world with our existing tools. And the models that we rely on often miss these critical details. I started to understand that models are only as good as their input data.

Property risk is generally described as a combination of hazard, exposure, and vulnerability. Of the three, vulnerability has received far less attention because it is far harder to observe: What is the roof made of? How much combustible vegetation surrounds the structure? Has mitigation occurred? How has the property changed since the last inspection? What physical characteristics make one home survive while another burns?

Those questions sound deceptively simple. But answering them across millions of properties has historically been expensive, slow, and often impossible.

A breakthrough appears

Around this time, Scott and I were introduced to Dr. Ophir Almog. Ophir had spent decades in remote sensing and Earth observation, including serving as Chief Science Officer of Israel’s elite Unit 9900 before earning a PhD in Remote Sensing from the Technion. He approached imagery from a fundamentally different perspective.

Most people look at a satellite image and see objects: a building, a road, a forest. Modern computer vision systems largely do the same, recognizing patterns and assigning labels. Ophir’s work started from a different question. Not, “what is it?” But “what is it made of?”

For decades, answering that question required hyperspectral sensors capable of measuring hundreds of wavelengths beyond human vision. These systems can reveal material composition, moisture content, vegetation health, and physical condition invisible in conventional imagery. They are also expensive, uncommon, and often tied to defense and intelligence applications.

The industry assumption was straightforward: if you don’t have hyperspectral imagery, you don’t get hyperspectral insights. Ophir spent years challenging that assumption. His work focused on spectral super-resolution, the reconstruction of hidden spectral information from ordinary imagery. In simple terms, recovering physical signals from data most people considered insufficient.

When we first encountered the work, we weren’t thinking about starting a company, only whether there might be a fundamentally better way to measure the physical world. The more we learned, the clearer it became that the real bottleneck wasn’t access to imagery. Commercial satellites already collect petabytes of Earth observation data every year.

The bottleneck is the ground truth. The condition and material intelligence that answer the questions posed across so many of my conversations. We have become very good at collecting pictures of the world. We remain surprisingly poor at understanding the physical reality behind those pictures.

That realization arrived as artificial intelligence was advancing at extraordinary speed. AI models could write, reason, code, and generate content, yet remained surprisingly limited in understanding the physical world. They could describe wildfire risk. But they could not reliably tell you whether a roof was asphalt or concrete tile, or whether mitigation had materially changed a property’s vulnerability.

Insurance is simply where the consequences are easiest to see. Premiums rise. Carriers exit markets. Communities become harder to insure. Capital retreats from uncertainty. But the same challenge exists across infrastructure, agriculture, energy, supply chains, environmental monitoring, and eventually what many are now calling physical AI.

Wildfire became our starting point because it sits at the intersection of these forces and we had been impacted by it firsthand. The losses are happening, the stakes are immediate, and regulators in California and Colorado are already pushing the industry toward more granular, property-level understanding of risk and mitigation.

We started this journey thinking we were studying climate risk. We eventually realized we were studying an intelligence problem. That’s why we started OmniGeo: to help insurance and finance underwrite, price, and monitor physical risk at the individual asset level.

Where we are today

For the past two years, we’ve been building largely out of sight, focused on a different question than the one dominating the AI conversation: how do you teach machines to measure the physical world? That work led us deep into spectral science, remote sensing, material intelligence, and what we believe is an entirely new category of geospatial reasoning technology - years spent validating the science, building the platform, and testing it with early partners.

Today, OmniGeo is coming out of stealth. We’re launching our first product with a few design partners, a property-level, wildfire risk API for P&C underwriters, risk providers and wildfire modelers called OmniFire.

The product converts satellite, aerial, and drone imagery into dynamic, physics-based property risk signals that help insurers price, underwrite and monitor wildfire risk at the parcel level. Rather than relying solely on static inspections or coarse regional models, carriers can continuously measure the physical characteristics that influence vulnerability, including building materials, defensible space, vegetation condition, fuel density, and property-level change over time.

We chose property wildfire risk as our first market because the need is immediate and the economic impact is measurable. Better measurement creates better underwriting, stronger incentives for mitigation, and ultimately more resilient communities. We are currently in beta with a small group of carriers and risk providers with portfolio exposure across California and Colorado.

The measurement layer behind Ground Truth

OmniGeo’s first product, OmniFire, closes the earth measurement gap for insurers in wildfire-prone areas. OmniFire turns standard imagery into parcel level measurements of what a property is actually made of, including roof material, defensible space, and fuel moisture. It’s vulnerability measured at scale and it’s live today. If you underwrite, price, or manage wildfire risk, visit www.omnigeo.ai/omnifire to learn more.