Data Driven Stochastic Models of Wildfire Spread
Conference Year
January 2020
Abstract
Across the globe, the frequency and size of wildfire events are increasing. Research focused on minimizing wildfire is critically needed to mitigate impending humanitarian and environmental crises. Real-time wildfire response is dependent on timely and accurate prediction of dynamic wildfire fronts. Current models used to inform decisions made by the U.S. Forest Service, such as Farsite, FlamMap and Behave do not incorporate modern remotely sensed wildfire records and are typically deterministic, making uncertainty calculations difficult. In this research, we tested two methods that combine artificial intelligence with remote sensing data. First, a stochastic cellular automata that learns algebraic expressions was fit to the spread of synthetic wildfire through symbolic regression. The validity of the genetic program was tested against synthetic spreading behavior driven by a balanced logistic model. We also tested a deep learning approach to wildfire fire perimeter prediction. Trained on a time-series of geolocated fire perimeters, atmospheric conditions, and satellite images, a deep convolutional neural network forecasts the evolution of the fire front in 24-hour intervals. The approach yielded several relevant high-level abstractions of input data such as NDVI vegetation indexes and produced promising initial results. These novel data-driven methods leveraged abundant and accessible remote sensing data, which are largely unused in industry level wildfire modelling. This work represents a step forward in wildfire modeling through a curated aggregation of satellite image spectral layers, historic wildfire perimeter maps, LIDAR, atmospheric conditions, and two novel simulation models. The results can be used to train and validate future wildfire models, and offer viable alternatives to current benchmark physics-based models used in industry.
Primary Faculty Mentor Name
Chris Danforth
Graduate Student Mentors
David Rushing Dewhurst
Faculty/Staff Collaborators
David Landay, Todd DeLuca, Karl Kaiser, Nat Shenton
Status
Graduate
Student College
College of Engineering and Mathematical Sciences
Program/Major
Complex Systems
Primary Research Category
Engineering & Physical Sciences
Data Driven Stochastic Models of Wildfire Spread
Across the globe, the frequency and size of wildfire events are increasing. Research focused on minimizing wildfire is critically needed to mitigate impending humanitarian and environmental crises. Real-time wildfire response is dependent on timely and accurate prediction of dynamic wildfire fronts. Current models used to inform decisions made by the U.S. Forest Service, such as Farsite, FlamMap and Behave do not incorporate modern remotely sensed wildfire records and are typically deterministic, making uncertainty calculations difficult. In this research, we tested two methods that combine artificial intelligence with remote sensing data. First, a stochastic cellular automata that learns algebraic expressions was fit to the spread of synthetic wildfire through symbolic regression. The validity of the genetic program was tested against synthetic spreading behavior driven by a balanced logistic model. We also tested a deep learning approach to wildfire fire perimeter prediction. Trained on a time-series of geolocated fire perimeters, atmospheric conditions, and satellite images, a deep convolutional neural network forecasts the evolution of the fire front in 24-hour intervals. The approach yielded several relevant high-level abstractions of input data such as NDVI vegetation indexes and produced promising initial results. These novel data-driven methods leveraged abundant and accessible remote sensing data, which are largely unused in industry level wildfire modelling. This work represents a step forward in wildfire modeling through a curated aggregation of satellite image spectral layers, historic wildfire perimeter maps, LIDAR, atmospheric conditions, and two novel simulation models. The results can be used to train and validate future wildfire models, and offer viable alternatives to current benchmark physics-based models used in industry.