Elsevier

Safety Science

Volume 139, July 2021, 105196
Safety Science

Determining the likelihood of asset destruction during wildfires: Modelling house destruction with fire simulator outputs and local-scale landscape properties

https://doi.org/10.1016/j.ssci.2021.105196Get rights and content

Highlights

  • The destruction of houses is a common impact of wildfires.

  • Fire simulations are often used to understand patterns of landscape fire risk.

  • Simulations of historic fires were used to develop a house destruction algorithm.

  • House loss can be predicted given accurate fire estimates and suitable training data.

  • Reconstructed fires can be used to add value to risk simulation systems.

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.

Introduction

Wildfires are a common occurrence on all vegetated continents on earth (Krawchuk et al., 2009). When they occur in populated areas, they have the potential to impose substantial impacts to infrastructure, property and lives (Tedim et al., 2018). Extreme wildfires that impact populations are highly disruptive and recovery of human systems can be protracted and expensive (Chang-Richards et al., 2013, Ambrey et al., 2017). Recent examples include the 2016 Fort McMurray fire in Canada (Kochtubajda et al., 2017), the Carr fire in California in 2018 (Lareau et al., 2018) and the fires in Portugal in 2017 (Brown et al., 2018). Worldwide, there is evidence of increasing wildfire danger and severity (Jolly et al., 2015, Abatzoglou and Williams, 2016, Singleton et al., 2019); furthermore, projections of climate indicate that the occurrence of the weather conditions that promote such fires are expected to become more frequent (Hennessy et al., 2005, Bedia et al., 2014, Wotton et al., 2017).

Land management agencies in susceptible areas to invest in preparedness activities to prevent or mitigate fire impacts (Berry et al., 2006, Milne et al., 2014). The costs of preparedness can be substantial, as they can include investment in suppression resources, fire detection systems, fuel management, public education, training regulation and the development robust infrastructure (Martell, 1982). A commonly considered metric of fire impact is the number of houses destroyed (Wilson and Ferguson, 1986, Blanchi et al., 2010, Gibbons et al., 2012, Penman et al., 2015, Gibbons et al., 2018). Dwellings have high societal value and the number of houses destroyed during a wildfire has been shown to be positively correlated with the number of human lives lost (Blanchi et al., 2012) – another commonly used metric of fire impact. Houses are commonly used as refuges during fires (Blanchi et al., 2014), and incidents in which large numbers of houses areburnt have been shown to have complex and long term impacts on the community at large (Kumagai et al., 2004, Kulig et al., 2013, Ambrey et al., 2017). To be able prioritise how to best allocate preparedness resources to reduce the risk of fire to houses requires an understanding of the consequences of fire and an understanding how preparedness and suppression activities may alter these (Thompson et al., 2017).

Houses can be ignited by direct flame contact, radiant heat or embers – either individually or in combination (Barrow, 1945, Ramsay et al., 1987, Beverly et al., 2010, Caton et al., 2017). Factors that influence house exposure to ignitions and consequent potential for destruction have been found to operate at a range of spatial scales. For example, at small scales (10′s of meters) concentrations of flammable vegetation and other fuels have been shown to increase likelihood of a house being burnt during a fire as they can increase the direct exposure of assets to ignition sources and reduce the ‘defensible space’ that allows fire suppression (Wilson and Ferguson, 1986, Gibbons et al., 2012, Syphard et al., 2014, Gibbons et al., 2018, Penman et al., 2018). Similarly, where houses are located in close proximity to each other, the ignition of one house may affect the likelihood of others being destroyed (Blanchi and Leonard, 2005, Cohen and Stratton, 2008, Alexandre et al., 2016a, Alexandre et al., 2016b).

Most studies have focused on the local-scale influences on ignitions, however the interaction of these with larger scale aspects of fire behaviour will ultimately determine fire exposure (Cheney, 1976, Wilson and Ferguson, 1986). Areas burnt by head fires are exposed to higher intensity combustion and, as they are windward of the fire, more firebrands (Koo et al., 2010). Furthermore, large areas of fuel burning at high intensity can become ‘plume driven’, exhibiting atmospheric coupling (Sun et al., 2009, Morvan and Frangieh, 2018) which is thought to further increase building vulnerability due to pyrogenic winds (Cheney, 1976, McRae et al., 2013). Large scale landscape properties that influence fire behaviour have been utilised to estimate the likelihood of house loss, including the effects of slope, heavy fuels and fuel connectivity (Price and Bradstock, 2013, Alexandre et al., 2016b). However, as fire behaviour is undeniably a function of the weather during a fire incident, approaches that estimate house risk without considering local, landscape and weather effects may not be able accurately quantify potential risk. The risk of any particular house in the landscape will be a function of the proximal position of the house relative to fuel, topography and other houses, and the likely conditions and fire behaviour during any wildfires that occur. Furthermore, there is evidence that factors that contribute to house destruction are not necessarily additive and there may be interactions. For example, the effectiveness of suppression can be greatly enhanced when the level of fire exposure is reduced (Hakes et al., 2017).

One way the elements that define the likelihood of house loss at different scales can be combined is fire behaviour simulation. Fire simulation systems enable users to simulate the spread and nature of hypothetical fires through the representations of the landscape (Perry, 1998). They are parameterised with information relating to fuel, topography, ignition locations and potential weather, and use these to create predictions of potential fire behaviour (Miller and Ager, 2013). The results of ensembles of simulations can be aggregated to allow users to identify parts of the landscape that are likely to be most vulnerable to frequent or extreme fire behaviour (Finney et al., 2011, Ager et al., 2013, Haas et al., 2013, Duff et al., 2014, Castillo et al., 2017). By overlaying simulation outputs and assets, the fire exposure can be estimated (Ager and Vaillant, 2011, Driscoll et al., 2016). Such approaches have been adopted operationally; for example, in the state of Victoria, Australia, ensembles of fire simulations using PHOENIX RapidFire (PHOENIX, (Tolhurst et al., 2008)) are used to identify the parts of the landscape most at risk to fire and provide an objective evaluation of preparedness (Department of Environment and Primary Industries, 2013). In this approach, to compare between scenarios, risk is quantified by finding the total number of houses to be expected to be destroyed. Expected destruction is determined by overlaying simulated outputs of fire behaviour over maps of houses and predicting destruction based using a linear logistic regression of fire behaviour variables that had been created using observations of destruction in four fires (Tolhurst and Chong, 2011).

To-date, there have been no approaches developed that facilitate the prediction of asset destruction that consider likely fire behaviour, local-scale land influences and the interactions between factors. We aim to develop and test a process that can be integrated with existing fire simulation-based risk evaluation systems to enhance the representation of fire impacts. We will demonstrate this by linking observations of historic house destruction within wildfires in South Eastern Australia to local-scale landscape properties and fire behaviour reconstructed using the fire simulator that is operationally implementated in the study area. The resultant model has the potential to be applied to the simulator outputs for hypothetical fires, enhancing the ability of land managers to estimate the likely effects on risk of differing management strategies. This could also allow the primary factors contributing to house loss to be identified in the context of real landscapes, allowing risk reduction efforts to be targeted on a case-by-case basis.

Specifically, we aim to:

  • -

    Use historic fire events to train a tree-based machine learning model to predict the probability of house loss using outputs of the fire simulator PHOENIX and local scale landscape properties;

  • -

    Compare the new model with the existing regression-based house destruction model used with PHOENIX; and

  • -

    Evaluate the model outputs to assess the factors in the landscape that play the greatest role in determining house destruction from simulated fires

Section snippets

Study area

The study considers large wildfires that occurred in the state of Victoria, South Eastern Australia. The climate is predominantly Oceanic Marine west coast (Cfb) according to the Koppen climate classification as delineated by Chen and Chen (2013), although average summer temperatures are higher than is typical for the class. Summers are hot and dry, and wildfires are most frequent in the summer and autumn months (Murphy et al., 2013). Fires occur every year, however damaging fires are limited

Results

The final simplified model BRT model produced consisted of 2,400 trees and was found to explain 55.3% of the variation in the likelihood of house destruction. A 20-fold cross validation test using the final model structure (trained on 95% of the data and predicting to the remaining 5%) found that 30.1% of the variation (SD 0.08%) of the independent data was explained. When converted to binary outcomes and compared to the training data, the ACC of the BRT for predicting destruction was 90.5%,

Discussion

There have been numerous studies that have evaluated the risk of building destruction by fire, however approaches are typically either focused on likely fire conditions or the landscape features in the vicinity of the assets of interest. Our results indicate that to accurately estimate risk, the consideration of both properties is important. The contributions of the fire-derived (including convection, winds and embers) and landscape derived (including house density, fuel and road proximity)

Conclusion

To best understand the potential for house loss, we have found that it is important to recognise the combined effects of landscape properties in the proximity of assets and likely conditions during fires. The approach demonstrated here is a way that these effects can be incorporated into a system for predicting likely asset destruction in a manner that can be implemented in existing fire simulator-based risk evaluation frameworks. Robust asset destruction models implemented within such

Data availability

The fire progression data and house loss data were used under license with the Victorian Department of Environment, Land, Water and Planning. All data will be archived on University servers.

Funding Source

This work was undertaken as part of the ‘integrated Forest Ecosystem Research (iFER)’ program funded by the Victorian Department of Environment, Land, Water and Planning (DELWP).

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work was undertaken as part of project integrated Forest Ecosystem Research (iFER) program funded by the Victorian Department of Environment, Land, Water and Planning. We would like to thank the staff of the Victorian Department of Environment, Land, Water and Planning for their contributions to this research, in particular Andy Ackland and Finley Roberts.

References (90)

  • M.P. Singleton et al.

    Increasing trends in high-severity fire in the southwestern USA from 1984 to 2015

    For. Ecol. Manage.

    (2019)
  • J.T. Abatzoglou et al.

    Impact of anthropogenic climate change on wildfire across western US forests

    PNAS

    (2016)
  • A.A. Ager et al.

    Wildfire exposure analysis on the national forests in the Pacific Northwest, USA

    Risk Anal.

    (2013)
  • A.A. Ager et al.

    Finney, MA (2011) Integrating fire behavior models and geospatial analysis for wildland fire risk assessment and fuel management planning

    J. Combust.

    (2011)
  • P.M. Alexandre et al.

    Factors related to building loss due to wildfires in the conterminous United States

    Ecol. Appl.

    (2016)
  • P.M. Alexandre et al.

    The relative impacts of vegetation, topography and spatial arrangement on building loss to wildfires in case studies of California and Colorado

    Landscape Ecol.

    (2016)
  • Ambrey, CL, Fleming, CM, Manning, M (2017) The social cost of the Black Saturday bushfires. Australian Journal of...
  • G.J. Barrow

    A survey of the houses affected in the Beaumaris fire

    J. Council Sci. Industrial Res.

    (1945)
  • J. Bedia et al.

    Forest fire danger projections in the Mediterranean using ENSEMBLES regional climate change scenarios

    Clim. Change

    (2014)
  • A.H. Berry et al.

    Prescribed burning costs and the WUI: economic effects in the Pacific Northwest

    West. J. Appl. For.

    (2006)
  • J.L. Beverly et al.

    Assessing the exposure of the built environment to potential ignition sources generated from vegetative fuel

    Int. J. Wildland Fire

    (2010)
  • R. Blanchi et al.

    Investigation of bushfire attack mechanisms resulting in house loss in the ACT bushfire 2003

    (2005)
  • R. Blanchi et al.

    Life and house loss database: description and analysis

    (2012)
  • R. Blanchi et al.

    Meteorological conditions and wildfire-related houseloss in Australia

    Int. J. Wildland Fire

    (2010)
  • Costermans, L., 2009. Native Trees and Shrubs of South-Eastern Australia. New...
  • S.E. Caton et al.

    Review of pathways for building fire spread in the wildland urban interface part I: Exposure conditions

    Fire Technol.

    (2017)
  • Y. Chang-Richards et al.

    Resource challenges for housing reconstruction

    Disaster Prevent. Manage.

    (2013)
  • N.P. Cheney

    Bushfire disasters in Australia, 1945–1975

    Australian Forest.

    (1976)
  • N.P. Cheney et al.

    Grassfires: Fuel, Weather and Fire Behaviour

    (1997)
  • D.M. Chong et al.

    PHOENIX RapidFire 4.0 Convection and Ember Dispersal Model

    Bushfire and Natural Hazards Cooperative Research Centre

    (2012)
  • J. Cohen

    The wildland-urban interface fire problem

    Fremontia

    (2010)
  • J.D. Cohen et al.

    Home destruction examination: Grass Valley fire

    USDA No.

    (2008)
  • P. Cunningham et al.

    Severe convective storms initiated by intense wildfires: Numerical simulations of pyro-convection and pyro-tornadogenesis

    Geophys. Res. Lett.

    (2009)
  • G. De'ath

    Boosted trees for ecological modelling and prediction

    Ecology

    (2007)
  • Department of Environment and Primary Industries (2013) Victorian bushfire risk profiles: a foundational framework for...
  • D.A. Driscoll et al.

    Resolving future fire management conflicts using multicriteria decision making

    Conserv. Biol.

    (2016)
  • Duff, TJ, Chong, DM, Cirulis, BA, Walsh, SF, Penman, TD, Tolhurst, KG (2014) Understanding risk: representing fire...
  • J. Elith et al.

    Novel methods improve prediction of species' distributions from occurrence data

    Ecography

    (2006)
  • J. Elith et al.

    A working guide to boosted regression trees

    J. Anim. Ecol.

    (2008)
  • M.A. Finney et al.

    A method for ensemble wildland fire simulation

    Environ. Model. Assess.

    (2011)
  • M.A. Finney et al.

    A review of fire interactions and mass fires

    J. Combust.

    (2011)
  • J.H. Friedman

    Greedy function approximation: A gradient boosting machine

    Ann. Statist.

    (2001)
  • N. Gellie et al.

    Reconstructing the spread and behaviour of the February 2009 Victorian Fires

    Proceedings of the 3rd Fire Behaviour and Fuels Conference. Spokane, Washington, United States

    (2010)
  • P. Gibbons et al.

    Land management practices associated with house loss in wildfires

    PLoS ONE

    (2012)
  • A.M. Gill et al.

    Scientific and social challenges for the management of fire-prone wildland–urban interfaces

    Environ. Res. Lett.

    (2009)
  • Cited by (8)

    • Integrated wildfire danger models and factors: A review

      2023, Science of the Total Environment
    • Satellite remote sensing of active fires: History and current status, applications and future requirements

      2021, Remote Sensing of Environment
      Citation Excerpt :

      Landscape fire is a widespread natural disturbance agent involved in Earth's biogeochemical cycling, but one that can be greatly influenced by human actions, including in relation to climate and environmental change. Fire provides multiple ecological benefits (McLauchlan et al., 2020) – but in certain circumstances also poses a risk to life and infrastructure (Duff and Penman, 2021). In areas of substantial landscape fire activity and in regions downwind, air quality can also be seriously degraded - leading to major human health impacts and hundreds of thousands of early deaths per year worldwide (Roberts and Wooster, 2021).

    • Bottlenecks, feedbacks and thresholds in wildfire suppression

      2023, Proceedings of the International Congress on Modelling and Simulation, MODSIM
    View all citing articles on Scopus
    View full text