An empirical machine learning method for predicting potential fire control locations for pre-fire planning and operational fire management

Abstract

During active fire incidents, decisions regarding where and how to safely and effectively deploy resources to meet management objectives are often made under rapidly evolving conditions, with limited time to assess management strategies or for development of backup plans if initial efforts prove unsuccessful. Under all but the most extreme fire weather conditions, topography and fuels are significant factors affecting potential fire spread and burn severity. We leverage these relationships to quantify the effects of topography, fuel characteristics, road networks and fire suppression effort on the perimeter locations of 238 large fires, and develop a predictive model of potential fire control locations spanning a range of fuel types, topographic features and natural and anthropogenic barriers to fire spread, on a 34000km2 landscape in southern Idaho and northern Nevada. The boosted logistic regression model correctly classified final fire perimeter locations on an independent dataset with 69% accuracy without consideration of weather conditions on individual fires. The resulting fire control probability surface has potential for reducing unnecessary exposure for fire responders, coordinating pre-fire planning for operational fire response, and as a network of locations to incorporate into spatial fire planning to better align fire operations with land management objectives.