Maggie Epstein is a second year PhD student in the Fire Center at University of Montana where she also received her master’s degree. She comes from a wildland fire fighting background and worked on a rappel crew in northern Idaho. Her research revolves around bridging the gaps between operational firefighting and fire science.
She helps SMART FIRES researchers put on prescribed fires so that other scienctists can attend and gain access to the situations and data they require. However, her research revolving around firefighter decision making and large language models is incredibly relevant to SMART FIRES. Please see her interview below.
What got you into research?
I loved research and academia as an undergrad and loved the process of finding questions and exploring them. I knew I was going to be a firefighter for my career path and initially grad school was what I wanted to do in off season. But I ended up taking it really far and falling in love with academia, too.
What was your master’s research?
It was in the biophysical modeling realm. I used machine learning on a combination of satellite sensors and lidar sensors to get really robust, detailed pictures of what the canopy looks like. This helps us understand how it might burn but I looked at the inverse question of that. I looked at how canopies recover after fire and essentially created detailed, time series of wilderness forests to see what they were doing post fire on a large scale, as in the entire Selway-Bitterroot Wilderness, Glacier National Park, and Bob Marshall Wilderness.
How has diving into the research of fire science changed your relationship to firefighting?
They are two different disciplines that, for the most part, operate in their own lanes. One of the big motivators when I started my PhD was finding avenues forthe knowledge and the on the ground observation of firefighters to work its way into academia. So I honestly think firefighting informs my experience as a researcher more than the other way.
What is your PhD project?
I work with a data set called the Wildland Fire Decision Support System. When we have really fig fires in the west that go on for weeks or months and have hundreds of firefighters on them, firefighters chart about their fires the way a doctor might chart about a patient. They document their decision making and why we are doing the things we’re doing.
We have these free text entries in this system for every big fire in the west – a little under 7,000 fires. And no one’s done anything with this information because its 100,000 pages of just people writing. You can’t do much with it using traditional scientific means but it is this wealth of information. So I trained a large language model to read through WFDs and identify strategic barriers on 6630 fires in the western US.
And really the punchline of that is that right now through bipartisan infrastructure law we’re spending a ton of time building field treatments, and there’s a lot of scientific thought going into how to construct effective field treatments but no one really knows if we actually use those strategically to fight fire.
Its also a little like a anthropological documentation of how we fight fire culturally.
What were the difficulties in using a Large Language Model?
We used a commercially available large language model that we fine-tuned, so the actual model was a bit of a black box to us. and that we didn't know the training set that was used on it or the exact model structure.
We know that large language models are neural nets for the most part, but we didn't know the structure of that particular one. So it's a lot of experimenting to figure out where the holes are in the training data. And it's also really, really literal. Firefighters have their own super specific language that can be a little bit gate kept and doesn't even always make sense to scientists.
It was a very iterative process for sure. And kind of the big caveat is we only use this model to like parse text and identify things that were in the text. And it's really quite good at that. I probably wouldn't use this model to like generate original text with any confidence.
What are the next steps in your research?
It is a big data set with all these facets of analysis and we just have to pick a direction to go. At some point we would really like to link it back to like actual physical landscape features. Right now our source of information are these big bodies attacks for firefighters which is where we talk about these things on the landscape, but we don't have anything measured on the landscape. I can see where a previous prescribed fire but I don’t know anything about it. I don’t know how much of an impact the burn made on biomass, how big it is, or how its oriented. We would really love to link these back to something measurable.