A poster by three Colorado College students has been accepted for presentation at the 22nd annual Posters on the Hill, to be held April 17-18 in Washington, D.C. Only 60 posters were accepted from the more than 400 applications received for the conference, which is sponsored by the Council on Undergraduate Research.
Anna Hessler ’18, a computer science and Spanish major from Denver, Nick Crews ’18, a computer science and physics major from Girdwood, Alaska, and David Radke ’18, a computer science major and discrete math minor from Orinda, California, will present “Using Artificial Neural Networks to Predict Wildfire Growth.”
In a letter to the organizers recommending the poster, Assistant Professor Matthew Whitehead and Visiting Professor Daniel Ellsworth, both with CC’s Department of Mathematics and Computer Science, noted that wildfires are extremely dangerous and costly. “The US Forest Service alone spent more than $2 billion on firefighting efforts in 2017,” the submission said. “The research done in ‘Using Artificial Neural Networks to Predict Wildfire Growth’ aims to substantially improve fire control efforts by providing better estimates for fire growth than currently available models. If successful, this work will help to reduce the time to contain wildfires, better predict where evacuations may be needed, and support improved simulations for use in training.”
The students’ model was trained on historical weather data, historical fire perimeters, digital elevation models, and historical satellite imagery. After evaluating this vast quantity of historical data, it is able to extract trends and patterns, so that given a new input of weather data and an updated state of a fire, it is able to predict the fire perimeter for the next day.
“We realized exactly how relevant this project could be when we saw the physical impacts of wild fires on communities, especially in recent years,” says Hessler.
Radke developed the poster idea after working on a research project at UC Berkeley last summer that focused on the effects that natural disasters could pose to California’s fuel transportation infrastructure (roads, pipelines, refineries, etc.) through 2100. “We used many different kinds of software on the project, including pre-existing fire modeling software,” Radke says. “Accuracy with some of these models was always in question, and with my senior capstone fast approaching, I decided I wanted to build a piece of software that could better model the spread of wildfires. After arriving on campus, I ran the idea by Nick and Anna, and after a couple days of brainstorming, we decided on this project. Using an artificial neural network to map wildfire is an approach that hasn’t really been explored, and we thought it would be a really cool project moving forward.”
A neural network is a mathematical model of how the brain works, with “neurons” connected to each other with “synapses,” says Crews. When one neuron is “stimulated” by an input, it stimulates all its neighbors. “By connecting them in a network, you can pass an input signal into one side of the network and an output signal comes out the other side,” he says. “They are useful because you can ‘train’ them, by feeding them good examples of input/output pairs; for example, what the weather/fire conditions were the day before, and what they were the following day, and the network is able to update itself so it can emulate those predictions.”
Helping with the project was Matt Cooney, GIS specialist in the newly renovated Tutt Library GIS lab. “He certainly helped a lot early on when we were needing to gather the remotely sensed data for all of our fire locations and format them so we could export them to our neural network,” says Crews.
The students also gathered historical weather data from the National Oceanic and Atmospheric Association, satellite imagery and Digital Elevation Models from the USGS National Map, and fire perimeters from the USGS GeoMAC database.
In the submission, Whitehead and Ellsworth also note that “machine learning techniques are an active and important research area in computer science. Deep artificial neural networks have been shown to have strong classification performance when working with large collections of scientific data and visual information. This project harnesses the power of convolutional neural networks to predict wildfire growth in an automated fashion. A successful model capable of accurately predicting the spread of wildfires would be an important result in applied machine learning research because of the novel dataset being tested along with the immediate practical applications.”