Predictive Understanding of Wildfire Ignitions across the Western United States

Wildfires have been increasingly impacting people and the environment in the western United States in recent decades. Prevention is one of the most effective fire mitigation strategies given that a majority of wildfires across the United States are ignited by human activities, many of which are potentially preventable. We used more than half a million wildfire records from 2000-2020, and a rigorous set of samples that represent absence of fire conditions, to develop machine learning and deep learning models of daily wildfire ignition probability across the western United States. We included 29 physical, biological, social, and administrative attributes in our predictive models to capture the spatial and temporal patterns of wildfire ignitions. Our results indicated that the commonly used Hypersampling technique–that generates spatially uniform samples to represent absence of fire–can effectively capture the widely different attributes associated with presence and absence of wildfire ignition, but fails to capture more nuanced, subtle distinctions between them, rendering ignition probability maps that indicate invariably high fire risk across the western United States during critical fire danger conditions. We found that stratified sampling of absence of wildfire ignition in proximity to the location and date of historical wildfires can better inform ignition models with accuracies exceeding 95% across diverse performance metrics. This augmented model not only is capable of developing fire risk maps with nuanced spatial patterns and can be adopted operationally to inform fire prevention strategies, but also enhances the understanding of wildfire ignition drivers.