Wildfire management is entering a new era. As climate change drives larger and more unpredictable fire seasons, the tools used to model and predict wildfires are being reinvented from the ground up. At the center of this transformation is a partnership between cutting-edge geospatial compute infrastructure and decades of ecological research expertise.
In a recent interview with Geo Week News, Ben Pruden, who leads the go-to-market team at Wherobots, and Kit O’Connor, a research ecologist with the US Forest Service’s Rocky Mountain Research Station, discussed how their collaboration is bringing AI-driven wildfire containment prediction from research into the operational realm.
From Annual Snapshots to Daily Forecasts
O’Connor’s work builds on a model introduced in 2017 called the Potential Control Locations (PCL) model; it’s the first model trained on empirical observations of where fires historically stopped. Using a 20-year stack of large wildfire data and the Landfire program, which accounts for every pixel across the entire United States, the model used machine learning to identify relatively better or worse holding features across the country.
“It was strategically very interesting and useful,” O’Connor explained. By 2021, the Forest Service had standardized the PCL model as a product delivered west-wide prior to fire season, giving every incident management team access to that year’s best control features.
However, the model had the limitation of being an annual product. It averaged across 20 years of data, producing a static snapshot rather than a dynamic forecast.
Starting in 2022, in collaboration with Microsoft Research and Development and a significant injection of compute power, O’Connor’s team asked the more ambitious question of if they could model the daily variability in containment potential.
That required reinventing the entire modeling system. The team spent roughly two years doing quality assurance on historical daily fire progression records, data that is limited and rife with quality challenges. They then used those records to modify the underlying PCL model with 72 hours of hourly weather changes, incorporating factors like wind gust, temperature, and vapor pressure deficit, AKA the same conditions firefighters consider on the ground.
The result is FireCon, a neural network built on top of the original machine learning algorithm that can predict up to 48 hours in advance where containment features are likely to emerge and where they will dissolve.
The Scaling Challenge
When the team began testing FireCon on individual fires in 2025, they received clear feedback: a single-fire view wasn’t enough. Fire managers needed a regional picture of what resources were in use, what was available, and where conditions were changing across the West.
They needed to run FireCon westward. That immediately broke their infrastructure, running the model at that scale would have taken weeks. That’s when the partnership with Wherobots became essential.
Rasterflow: Taking Infrastructure Off the Table
Wherobots, built on the open-source Apache Sedona project, is a distributed compute system designed for geospatial workloads. Its newest capability, Rasterflow, is specifically built to handle Earth observation workflows like FireCon, taking the infrastructure burden off of modeling teams entirely.
“Once you have your model that’s ready to go, we can onboard it into our compute environment, deploy it, and run it across as many machines as required,” Pruden explained. Rasterflow leverages a combination of traditional CPUs and GPUs, distributing compute across the appropriate infrastructure for each model’s requirements.
A critical pre-step involves preparing the data. Satellite and imagery products come in different formats, resolutions, and coordinate reference systems. Rasterflow harmonizes all of those layers into an analysis-ready data cube before the model even runs. For FireCon, that means normalizing multiple inputs to the standard 30-meter resolution and then exporting predictions at a 250-meter resolution for the final control point outputs.
“All that heavy lifting—the raster datasets, the reprojection, the coordinate reference system management, all the way through to preparing the final analytical-ready data cube, running the model on that data cube, and delivering the outputs—that’s what we take care of,” Pruden said.
Looking ahead, Wherobots’ GA release of Rasterflow, coming next month, will make all of those steps not only runnable by a human but also triggerable and managed by a tasking API (Application Programming Interface) designed to be agent-first.
Bridging Coarse Weather Data and Fine-Grained Predictions
O’Connor added an important clarification about the data pipeline: the weather data, which is high-frequency, comes in at a coarse 3-kilometer resolution. That has to be downscaled all the way to 30 meters to match the PCL model, and the final product is also produced at that 30-meter standard.
“The reason for that is that fire behavior modeling on individual fires for about the last 16 years has used that 30-meter standard,” O’Connor said. “We want to make sure that any product we’re producing across the entire country is hitting that standard of what fire managers are used to using, because it’s familiar to them.”
This attention to usability is a core part of the partnership, which is pairing Wherobots’ technology with the expertise of management agencies that know how to put a product into the hands of people who can actually use it.
Checking Accuracy in a Messy World
The team has multiple methods for validating FireCon’s predictions. One key source is NIROPS—an infrared dataset originally used to train the model that is still collected daily on large fires. This provides an independent evaluation against which the team can compare predicted fire growth or dormant periods.
They can also summarize FireCon results by specific firefighting features: existing containment lines, Potential Operational Delineations (PODs), and current fire perimeters. This allows them to ask and answer whether a fire is expected to grow on a given day and whether it is likely to hold at any pre-identified control feature.
Still, O’Connor is clear-eyed about the model’s limitations. “On every given fire, there are many, many human factors that would influence success or failure for containment,” factors like how many resources are deployed and microclimate conditions that the model can’t fully account for.
The Future of Earth Observation AI in Wildfire Management
O’Connor sees Earth observation AI as massively accelerating the ability to model and understand wildfire events. Rapid-refresh, high-spatial-resolution data creates better opportunities to calibrate models and learn from past datasets, producing powerful predictive capabilities.
The ultimate goal is not just reactive but proactive: improving the safety profile for the public and for fire professionals and creating opportunities to manage wildfire more effectively, including expanding prescribed burning and other proactive land management strategies.
For Wherobots, the motivation is similarly fundamental. “We want to enable researchers, scientists, organizations, and businesses to analyze and understand and create data products that answer critical physical-world questions without having to worry about all the trouble of managing and building the infrastructure to run it,” Pruden said. “That’s really the bottleneck that we see many organizations face.”
By removing that bottleneck, Wherobots and the Forest Service are demonstrating what becomes possible when domain expertise and scalable compute infrastructure meet: a future where the people closest to the problem can focus on the science, not the servers.
