GC33E-0572:
Building a model to predict megafires using a machine learning approach.

Wednesday, 17 December 2014
Harry Richard Podschwit, University of Washington Seattle Campus, Seattle, WA, United States, Renaud Barbero, University of Idaho, Moscow, ID, United States, Narasimhan K Larkin, US Forest Service Research, PNW AirFire Team, Seattle, WA, United States and Ashley Steel, US Forest Service Seattle, Seattle, WA, United States
Abstract:
Weather and climate are critical influences of wildland fire activity. Climate change has led to an increase in the size and frequency of wildfires in many parts of the United States. These changes are expected to increase under current climate change scenarios, likely exacerbating so called “mega-fire” activity. Megafires are typically the most devastating and costly to suppress. It is then desirable to know when and where weather conditions will be conducive to the development of these fires in the future. However, standard statistical methods may not be suited to handle the data imbalance and high-dimensional features of such an analysis. We use an ensemble machine learning approach to estimate the risk of megafires based on weather and climate variables for each ecosystem in the contiguous U.S. Bootstrap aggregated trees are used to describe which suite of coarse scale weather conditions has historically best separated megafires from other large fires and to estimate the conditional probability of a “megafire” given ignition. The annual distribution of ignitions was estimated to calculate an overall probability of a “megafire” and spatial wildfire patterns were used to appropriately distribute this probability across space. This framework was then applied to future climate projections under the RCP8.5 scenario to estimate the future risk of these fire types. Our methodology was applied to various climate change scenarios and suggests that the frequency of these types of fires is likely to increase throughout much of the western United States in the next 50 years.