Modeling burn probability patterns for large fires
Date
2013
Authors
Ziesler, Pamela Sue, author
Rideout, Douglas B., advisor
Reich, Robin, committee member
Wei, Yu, committee member
Kling, Robert, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
I present a set of techniques for modeling burn probability patterns for large wildland fires. The resulting models address an important goal of a large fire risk analysis by estimating large fire burn probabilities. The intent was to develop models of burn probability using data that are widely available or easily calculated and that achieve acceptable predictive performance. Two models were successfully estimated using variables that may be extracted directly or easily calculated from standard GIS layers and other sources and they had `good' predictive ability with AUCs of 0.81 and 0.83. The ultimate intended use for the models is strategic program planning when information about future fire weather and event durations is unavailable and estimates of the average probabilistic shape and extent of large fires on a landscape are needed. Four primary objectives were to: estimate models from historical fire data that are appropriate for strategic program planning, incorporate the effect of barriers to the spread of fires across a landscape, account for the average effect of weather streams and management actions on large fires without using detailed information on weather, fire duration or management tactics, and investigate methods for addressing the spreading, connected nature of large fires on a landscape within a regression model. Models like these can provide finer detail than most landscape-wide models of burn probability and they have advantages over simulation methods because they do not require multiple runs of spread simulation models or information on fire duration or hourly weather events. To model burn probability patterns, I organized historical fire data from Yellowstone National Park, U.S.A. into a set of grids; one grid per fire. I incorporated explanatory variables such as fuel type, topography data and fire season indicators and I captured some spatial relationships through the use of distance, direction and other geometric variables. The data set observations are highly correlated and I investigated two approaches to account for and incorporate this correlation: one employed an autoregressive covariance structure and the other utilized a variable to account for the effects that neighboring cells may have on average burn probability. The two approaches yielded models with estimated coefficients that are consistent with fire behavior theory and that reflect how fires usually behave on the study site landscape. Both models compared well with the predictive ability of other fire probability models in the literature. Based on their predictive performance, this was a successful first attempt at addressing the research objectives and for estimating regression models to predict burn probability patterns for large fires.