Abstract
Wildfire spread is a stochastic phenomenon driven by a multitude of geophysical and anthropogenic factors. In this study, we propose a spatiotemporal data-driven risk assessment framework to understand the effect of various geophysical/anthropogenic factors on wildfire size, leveraging a systematic machine learning approach. We apply this framework in the state of California–the most vulnerable US state to wildfires. Using county-level annual wildfire data from 2001–2015, and various geophysical (e.g., land cover, wind, surface temperature) and anthropogenic features (e.g., population density, housing type), we trained, tested, and validated a suite of ensemble tree-based learning algorithms to identify and evaluate the key factors associated with wildfire size. The Extreme Gradient Boosting (XGBoost) algorithm outperformed all the other models in terms of generalization performance, categorization of important features, and risk performance. We found that standard deviations of meteorological variables with long-tailed distributions play a key role in predicting wildfire size. Specifically, the top ten factors associated with high risk of larger wildfires include larger standard deviations of surface temperature and vapor pressure deficit, higher wind gust, more grassy and barren land covers, lower night-time boundary layer height and higher population density. Our proposed risk assessment framework will help federal/state decision-makers to adequately plan for wildfire risk mitigation and resource allocation strategies.
Highlights
Wildfire is a highly dynamic phenomenon comprising of various physical interactions at different spatiotemporal scales
We provide an overview of the various data pre-processing steps leveraged to obtain the final dataset used for consecutive analysis, i.e., predicting the wildfire size and evaluating the key anthropogenic and geophysical factors that increase the risk of wildfire burnt acres
A data-driven spatiotemporal wildfire risk assessment approach is introduced in this study to help policy makers capturing county-level wildfire spread behavior, and identifying the associated key contributing factors
Summary
Wildfire is a highly dynamic phenomenon comprising of various physical interactions at different spatiotemporal scales. There exist sufficiently-detailed, accurate simulation tools that leverage numerical models like the Coupled Atmosphere-Wildland Fire Environment (CAW F E) model (Coen, 2013; Coen et al, 2020) and W RF − SF IRE or SF IRE (Spread FIRE) model (Mandel et al, 2011, 2014) to simulate the spread of a single wildfire incident These models are useful in understanding behavior of specific wildfire events, they cannot be used in studying wildfire spread patterns in a large area (e.g., a state, a country, or a continent) due to prohibitive computational cost. Models like the large-fire simulator system (Finney et al, 2011) are used to investigate wildfire spread in larger areas Such models comprise of simplified equations and assume fire-driving variables spatially or temporally invariant to keep the computational costs manageable. These models run comparatively faster; stochastic studies can be conducted and probabilistic maps can be produced leveraging such models (Thompson et al, 2011)
Sign-in/Register to view highlights & summary 
Submitted Version (
Free)
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call