Citation

Duff TJ, Penman TD. Safety Sci. 2021; 139: e105196.

Copyright

(Copyright © 2021, Elsevier Publishing)

DOI

10.1016/j.ssci.2021.105196

PMID

unavailable

Abstract

Preventing the loss of human lives and the destruction of houses is a focus of land management in fire prone environments. An understanding of the location and nature of risks to assets is important to plan mitigation activities. House destruction during wildfires has been found to be a function of landscape properties, including fuel proximity, topography and proximity to other houses. Additionally, conditions during wildfires are important, with houses likely to be destroyed when flames are intense. We propose an approach to predict destruction that considers both landscape attributes and fire conditions by using historic fires to train a destruction function. This was achieved by linking observations of house impacts in historic fires to fire behaviour reconstructed using a fire behaviour simulator. These observations were used to create an empirically fitted destruction function particular to the simulator used. We collated data for 17 wildfires that occurred in Victoria, Australia. These were replicated within the simulator PHOENIX RapidFire. The landscape data and simulator outputs were superimposed over house destruction/survival information and these were used to predict destruction likelihood. The best model explained 55% of the variation in house impacts, with fire and landscape predictors contributing to 50% of explained variation. It had an accuracy of 90.4%, in comparison to 46.4% of an alternative model that used simulated fire data alone. Incorporating empirically fitted asset impact functions in fire simulation systems can provide for improved projections of likely risk, supporting land managers to make evidence-based decisions that optimise the allocation of resources.
Summary
A key focus of wildfire management is the reduction of risk to life and property. We present an approach whereby observations of over 15,000 houses exposed to wildfires were used to develop models of the likelihood of destruction using a combination of simulated fire behaviour and landscape attributes. House destruction was found to be a function of both, with each contributing to around 50% of the model's performance, which accounted for 55% of observed variation. This approach can provide an indication of the possible processes that result in asset destruction and can be incorporated into fire simulation systems to better estimate potential impacts.

Language: en

Keywords

Asset; Boosted Regression Trees; Bushfire; Cost; Fire impact; Modeling; PHOENIX RapidFire; Risk; Wildland