Learning from the past, applying new tools to save firefighter lives
By Theresa Barosh | April 2025
Listen to the end to catch CIRA Director Steven Miller’s story about how the AI product GREMLIN came into existence.
Summary: In 2013, a wind burst from a thunderstorm led to the deaths of 19 wildland firefighters at the Yarnell Hill fire in Arizona. To support future firefighters, CIRA researchers developed a new tool using artificial intelligence to analyze satellite data and predict how storms will impact wildfire behavior. In 2021, the National Weather Center used the tool, GREMLIN, to successfully protect firefighters at a wildland fire outside Midland, Texas. GREMLIN stands for GOES Radar Estimation via Machine Learning to Inform NWP.
What is needed: continued funding and support for CIRA research combining satellite imagery and AI to save lives.
Full transcript:
Fire in the west is a growing problem with decades upon decades of drought. We are seeing major fire after major fire.
Over the recent years, I think the entire nation has started to appreciate the importance of understanding fire because smoke transports across the country and affects us here in Colorado quite a bit, but on the East Coast, we’ve seen plenty of major smoke events now as well. CIRA has been heavily involved with developing new tools for characterization of fire. So, monitoring a fire from space, for example, we have satellites up there right now that do a great job of detecting the fire, either in high spatial resolution and, in other cases, high temporal resolution so you can watch the fire evolve. And, of course, we have imagery tools that allow us to see the smoke being produced by these fires. And then that smoke being transported downwind, which affects air quality and our respiratory health.
One important aspect of that work is early detection of fires, where hopefully with the right near real-time tools you can spot a fire in its early stages and maybe go do something about it before it turns into one of these incredibly out of control and wild blazes. That’s something that we refer to right now as the Next Generation Fire System. It’s a collaboration with NOAA and other partner institutes to take satellite information and get it into a near real-time kind of alert framework, where a fire happens and boom, there’s an announcement of here’s where it is, and here’s its properties right now. Go do something, right? So that’s kind of the detection and early warning aspect of our activity.
Another one is the Fire Weather Testbed, where it’s kind of a repository of resources that is being managed by NOAA, both in terms of detection of fires, but also modeling. And what is a fire going to do next? Given that it’s here, here’s the fuel types. Here’s the weather. Here’s the winds and the moisture and the environment. Taking those things, all together and then trying to predict where that fire is going to go over the next several hours. So imagine a fire team or a hot shot crew out there on a wildfire, and they have a fire boss that’s telling them where to go to most effectively fight the fire, and you might have an incident meteorologist on site as well, who’s the weather liaison for what’s going to happen. These folks all need accurate and timely information on what that fire is doing and what we think it’s going to do next. So we’re plugged in with that aspect as well, which is a very kind of new and exciting way of approaching fire, where we’re consolidating all the resources together for a common good.
An interesting story that I like to tell about how CIRA takes problems and challenges and tries to address them into practical solutions has to do with a tragedy that happened in 2013 in Prescott, Arizona, where the Yarnell Hill fire was burning out of control, a wildfire. The Granite Mountain Hot Shots were a crew that were fighting this fire, and they were on, I believe, the southwestern line of that fire. Attacking the line on its blaze. And out of nowhere, kind of, convection – thunderstorms – built up in the east. And in that part of the. Country, the lower atmosphere is often very dry in the desert Southwest. So when rain falls into a dry atmosphere, it will evaporate, and that evaporation cools the air, and it creates what’s called negatively buoyant air. It sinks really quickly when it’s colder than its environment because of the evaporation of the rain, and it hits the surface, and it fans out almost like a pancake on a griddle. That air goes out in all directions. It’s called a cold pool or an outflow from the thunderstorm.
Well, in this case where the Granite Mountain Hot Shot crew were fighting the fire, this outflow came in from the east northeast, and it basically changed the entire direction and intensity of the fire, as the strong winds blew over from a completely different direction than what was blowing before. It caught the entire crew off guard, and they scrambled to get out of the way, but they could not. They had to put their fire blankets down and the fire raged over them, and very unfortunately, it killed all 17 of the crew. [correction: all 19 crew]
So, with that as a backdrop, we asked, you know, how could we possibly avoid this? What went wrong, you know? How could we avoid this in the future? And there was a realization that you know the models can do pretty good with convection and these thunderstorms that cause these outflows, but not always in the exact locations where and when and timing of where they’re actually happening in nature. And in this area, there’s a sparse radar coverage and terrain blockages that make the surface radars less useful.
So what we ended up doing was developing a satellite-based estimate of radar. And this is where machine learning came into play, where we took satellite information and lightning information and trained it against actual radar data to come up with a synthetic radar product. The name of that product is GREMLIN. I’m not going to try to spell that out – with using GOES and lightning, and that’s what the G and L are in there. With GREMLIN, we are actually able to render a synthetic radar picture. And, we found that taking that radar information and putting it into a model initialization – this is where you start the model off on an initial state so that it can run forward and see what’s going to happen next – using Gremlin and initializing that model we could actually do better with predicting where convection was happening than the actual radar data itself. So, we’re taking advantage in that case of satellite information that’s providing coverage in these areas where the radars are not providing us good coverage. These kinds of tools are really important because, you know, nature and history sometimes repeats itself in different ways.
Many years later, there was a fire in Midland, Texas – Southwest Texas, a grass fire. Firefighters were once again fighting on the Southern flank of that blaze. Once again, strong thunderstorms were forming on the northern area of the region, creating these outflow boundaries. Once again, the cold pools were racing down towards that fire, and it almost looked like history repeating itself. But in this case, the Midland Texas Weather Forecast Office, the [NOAA] National Weather Service were able to use satellite imagery in this case to actually say, hey, look, we see an outflow coming. And get your crew and the equipment off of that southern flank of that fire before that outflow arrives, and they were actually able to, in real time, receive that information, move off the southern flank, and outside of harm’s way.
So this is, you know, a kind of a vignette of the way that we function and what’s important for us is learning from the past, applying science and research toward new tools that can help avoid repeating history in the future.