Developing an impact index for the Australian Fire Danger Rating System: predicting potential structure loss from wildfires

Introduction

Wildfires and their potential impacts on life and property are becoming increasingly complex and uncertain. This can be attributed in part to the interactions between fire behaviour, increasing populations at the wildfire interface, variations in land-use, community preparedness levels (Whittaker and Clark 2021) and awareness and response to warnings (Strahan 2020; Whittaker et al. 2020). Globally, changes in land-use and climate processes have altered fire regimes, trending toward larger and higher-severity fires (Nolan et al. 2022) with enhanced overnight fire activity and multi-day fire events (Peace and McCaw 2024). Despite this increasing complexity and uncertainty, developing accurate predictions of the future impacts of wildfires remains a high priority for fire and land management agencies.

House loss by wildfire has been associated with the loss of life (Blanchi et al. 2012, 2014), with most fatalities associated with wildfires in Australia occurring among people evacuating from, sheltering in, or defending houses (Haynes et al. 2010). There are clear geographical trends in fatalities around population centres in south-eastern Australia, suggesting population density and/or housing density as key factors (Haynes et al. 2010; Blanchi et al. 2014). Accurately predicting house losses on a scale that informs decision-making for preparedness and planning to reduce house loss and improve human safety is needed; many current models predicting house loss have been developed at a resolution of 10s of metres and focus on localised or fine scale (individual structure) dynamics (Blanchi et al. 2006; Opie et al. 2014, Gibbons et al. 2018). These models have primarily been developed using large-scale events associated with large numbers of house losses under extreme fire weather behaviour, such as Ash Wednesday and Black Saturday in Victoria (Wilson and Ferguson 1986; Harris et al. 2011, 2012; Kilinc et al. 2013; Duff et al. 2019).

Other approaches have made use of successive runs of fire simulation programs, intended to delineate areas where fire and built-up areas are predicted to intersect (Massada et al. 2009; Ager et al. 2021). Deviations from normal weather patterns or heightened fire activity have also been used in actuarial models to model risk aggregated over large areas (Raveendran et al. 2024). Generally, studies are conducted at a fine spatial scale (metres), which is suited for risk planning (Tolhurst et al. 2008), but less applicable at the landscape level (kilometres). Models have highlighted fire behaviour characteristics such as fireline intensity and sudden changes in fire behaviour, in particular wind direction changes, as the core drivers of house loss (Ramsay et al. 1987; Leonard and Blanchi 2005; Blanchi et al. 2012; Harris et al. 2012; Kilinc et al. 2014; Tedim et al. 2018). The position of houses in the landscape, including proximity to fuel (Wilson and Ferguson 1986; Gibbons et al. 2012; Blanchi et al. 2014), surrounding fuel types (Kilinc et al. 2013; Collins et al. 2016; Penman et al. 2019), topography (Price et al. 2021) and housing density (Gibbons et al. 2012; Kilinc et al. 2013; Price and Bradstock 2013) have also been identified as important factors.

The destruction and damage of assets by fire are recorded for major fire events and some smaller events across Australia. Building Impact Assessments (BIA) conducted post fire by fire and land management agencies document the extent of damage to buildings (e.g. untouched, damaged or destroyed), building design and characteristics of the immediate surrounds. At a national scale, the spatial context of asset locations is important, as is the recognition that the density of assets influences the scale of potential loss. There is considerable diversity in the types of assets that could be impacted including houses, industrial buildings, critical infrastructure, cultural assets and natural resources. The different vulnerabilities associated with varied asset types and their environment influence community messaging and actions. For example, researchers (Harris et al. 2012; Kilinc et al. 2013) pointed out the destructive potential of fires depends not only on fire intensity but also on the size of the fire, suppression efforts and susceptibility to loss. The interplay of these factors and their attributes contribute to the difficulty in making credible forecasts of wildfire impacts.

The Australian Fire Danger Rating System (AFDRS; Matthews et al. 2019; Hollis et al. 2024), used by all Australian fire and land management agencies, improves on previous fire danger indices used in Australia by using current published fire behaviour models, regularly updated fuel maps and weather forecasts from the Bureau of Meteorology. The modular system employs one of eight fire behaviour models specific to the dominant fuel in each 1.5 × 1.5 km grid cell to estimate fire behaviour metrics such as the rate of spread, fireline intensity, flame height and spotting distance should a fire occur in a cell (Kenny et al. 2024). This data fills the resolution gap between local (structure-level) and district (fire weather area-level) data. The intent of the AFDRS is to provide fire and land managers with better decision-making tools and ultimately reduce the costs (social, economic and otherwise) associated with wildfire impacts.

In this study, our aim was to extend the predictive capability of the AFDRS to include the likelihood of structure loss in the landscape if a fire were to occur within, or spread to, a given cell. These predictions will be at a scale relevant to the AFDRS 1.5 × 1.5 km grid, allowing large-scale planning which is particularly important during fire seasons. The rapid identification of at-risk grid cells, especially when numerous active fires are present, makes the assimilation of relevant information manageable for planning, operations and community messaging teams. As a result, highlighted assets are less likely to be overlooked and targeted community communications are possible. The methods used to model the dataset are tailored specifically to discriminate between structures lost in wildfire and closely adjacent structures which were not lost (up to 250 m from structures lost), for application in the bushland–urban interface where structure losses tend to occur. Two modelling approaches were taken, modelling the loss of individual structures (binomial loss model), and another to predict proportional loss of structures (proportional loss model).

Outputs from this modelling will be interpreted alongside the existing AFDRS (Matthews et al. 2019) fire behaviour outputs (rate of spread, fireline intensity, flame height, spotting distance) which incorporate fire weather and generates the fire behaviour index (FBI). To not create a redundant index, as structure losses overwhelmingly occur during heightened fire behaviour and would give similar outputs to FBI if weather inputs were included, this modelling focusses on landscape and built-environment drivers of structure loss, rather than fire weather which is captured in the FBI. An important constraint of this work was that the predictors used in this model are required to be available at national scale and the models do not include assumptions about future events (e.g. fire direction, fire agency or homeowner fire suppression), as these would not be available as inputs in a useful timeframe, or would be computationally prohibitive to implement operationally.

Methods

Dataset

Data for structure loss during wildfire was provided by the New South Wales Rural Fire Service, Country Fire Authority Victoria, Country Fire Service South Australia and Queensland Fire and Emergency Services (n = 10,865, observed between February 2009 to March 2020). This included data collected in New South Wales (8296 records), Victoria (2358 records), Queensland (180 records) and South Australia (31 records). Structures (all buildings including houses, sheds and constructed shelters used for habitation, storage and other purposes) were addressed rather than using houses as the input data (described below), as it does not reliably discriminate between houses and other large structures.

Structure loss is defined here as complete destruction, whereas structures recorded as damaged (not destroyed) by the fire were excluded in the analysis due to the uncertainty around levels of damage and data accuracy. Losses were spatially matched to Microsoft Building Footprints centroids (https://github.com/microsoft/AustraliaBuildingFootprints) by assuming that structure loss coordinates within 50 m of the Microsoft dataset corresponded to the loss location. Where multiple candidate matches were returned (i.e. multiple losses within 50 m of a given point in the Microsoft data), the closest was assumed to be the correct match. Loss data coordinates with no match in the Microsoft data (>50 m from Microsoft coordinates, or displaced by closer loss coordinates), were assumed to be accurate, i.e. that the fire agency data are correct. The remaining unmatched Microsoft coordinates were then considered non-losses.

Data in the Microsoft Building Footprints was collected between 2013 and 2018, with no indication as to when an individual building was added to their dataset. This may lead to inaccuracy in structures which were built after a given fires’ occurrence being included as a non-loss, or structures built after the 2018 collection being absent where analysis for post 2018 fires were performed. Of the losses, 2118 occurred before 2013 (predominantly during the Black Saturday fires), and 8599 post-2018, with the balance (148) occurring between 2013 and 2018. Non-losses within 250 m of a loss were then extracted for further analysis (Fig. 1a), as these represented structures adjacent to destroyed structures which were not themselves destroyed. This choice of comparison reflects the need for the model to predict likelihood of loss for structures with similar risk profiles, rather than selection of a comparison group further removed from structures which were destroyed. The latter approach, while likely simpler to model would give less useful outputs, given the need to make predictions at the bushland–urban interface where fire risk to built assets is greatest. The final dataset comprised 10,865 structures lost in wildfire, and 11,971 adjacent non-losses.

Fig. 1.

Plots showing examples of the method to find neighbouring structures within a 250 m radius (blue points within, grey squares outside) to a structure loss (red triangle; (a)), extraction of the grid values (showing cleared land in white as an example) for a given structure (red triangle) at the radii defined in the Methods section (b), define the adjacency networks (ANs), with grey shading showing structures within the AN threshold distance (shown as a blue circle in (c)), and an example of how the ANs can relate to topography (d, colours indicate different AN gorups). Different point colours and shapes show the ANs in (d), with topography shown as shading (darker for higher ground).

Structures were then clustered into 1.5 km adjacency networks (ANs; 809 total) prior to analysis (Fig. 1c), where an individual network comprised structures within 1.5 km from one another, and any structure more than 1.5 km from the network was deemed to be in a separate network. Spatially grouping structures allowed for control of spatial autocorrelation effects in subsequent analyses. These effects include individual fires, suppression responses across different parts of the landscape and other spatially correlated features not captured by the predictor variables (Fig. 1c, d). The 1.5 km distance threshold was selected based on the spatial distribution of losses, ensuring it was neither too small (e.g. 200 m), which would result in overly fine groupings of structures, nor too large (e.g. 20 km), which could merge different fires into a single AN grouping.

Spatial predictors were extracted for each structure (both loss and non-loss cases), including the numbers of adjacent structures, proportion of cleared land, median canopy height (hereafter canopy height) and terrain ruggedness index (TRI). Data sources are shown in Table 1. Cleared land was defined based on AFDRS fuel grid classifications (urban, non-combustible, built-up, rural, pasture, grass, horticulture or crops) and converted to a binary format (1 = cleared, 0 = vegetated). Other predictors (distance to nearest vegetation, vegetation type, time since fire, etc., further details can be found in Supplementary Material S1) were initially tested but not included in the final modelling as they did not improve predictive performance. Structure counts, cleared land proportion and median canopy height were extracted within a radius of 50, 100, 150, 200, 250, 300, 400, 500, 750, and 1000 m of the structure coordinates (Fig. 1b) following the approach of Price and Bradstock (2013). TRI was also extracted for each point using the method of Wilson et al. (2007) to determine cell-wise TRI values using the eight adjoining grid cells.

Table 1.Spatial data sources used in the analysis, and which predictors were drawn from them.

Citation	Data	Extracted predictor	Predictor name
Microsoft building footprints	Centroids of structures	Counts of structures within a given radius.	struc250m
			struc750m
Matthews et al. (2019)	50 m grid fuel layer	Proportion of cleared land within a given radius. Area-weighted by intersection with the radius.	clear100m
			clear400m
Gallant et al. (2009)	~90 m grid elevation data	Terrain ruggedness index (TRI): the summed absolute difference in elevation between the cell of interest and the immediately adjacent cells.	TRI
Scarth et al. (2019), Scarth et al. (2023)	95% canopy interception height	Median canopy height (in m) within a given radius. N.B. data source gives height in decimetres.	canheight50
			canheight400

Predictor names used in the model equations are also shown.

Extraction of predictors at increasing distance from structures was used for two reasons: firstly the direction of attack from a fire was unknown and could not be accurately inferred from available data, and secondly this approach allowed us to test for the most informative radius for a given predictor, enabling us to determine the distance from a structure where each predictor has the most influence on structure loss, given the resolution of the available data.

Results

In determination of the most appropriate buffer distance, univariate models comparing the buffers from 0 to 1000 m showed the lowest BIC for the 250 and 750 m radius buffer for structure counts (Fig. 2a, c) for the binomial and proportional models respectively (the lower BIC, the greater explanatory value). Odds ratios showed higher structure counts were associated with lower probability of loss/lower proportion of loss (Fig. 2b, d, odds ratios <1). Cleared land at 400 m (binomial model) and 100 m (proportional model) were most informative (Fig. 2e, g), with greater proportions of cleared land related to lower probability/proportion of loss (Fig. 2f, h). Canopy height at the 400 m radii (binomial model; Fig. 2i) and 50 m radii (proportional model; Fig. 2k) returned the lowest BIC. Taller canopy height was associated with increased probability of loss/proportional loss (Fig. 2j, l). Equations for the multivariate models built with these predictors and the inclusion of TRI are shown below:

bin_lp (binomial) and prop_lp (proportional) can then be transformed to an estimate of structure loss probability or proportional loss using:

where:

Fig. 2.

Plots of the ΔBayesian Information Criterion (BIC) (relative to the most informative BIC) of the univariate models fitted to each buffered value (left side, plots a, c, e, g, i, and k) and odds ratios (right side, plots b, d, f, h, j, l), for structure counts (1st and 2nd rows; ln transformed), cleared land proportion (3rd and 4th rows; logit transformed), and canopy height (5th and 6th rows; ln transformed). The 1st, 3rd and 5th rows show results from the binomial model, and the remainder the proportional model. In all plots, the lowest (best) ΔBIC predictor is indicated by a red rhombus point, with blue circles used elsewhere.

In both models the effects of the shared predictors showed similar relationships. Structure count was the most explanatory predictor in both the binomial and proportional models, as indicated by the largest BIC increase upon removal (Table 2), followed by TRI, canopy height and proportion of cleared land in decreasing importance. Higher structure counts were significantly associated with lower probability of loss at an individual structure level (Fig. 3a), and lower proportional loss (Fig. 3b). Greater cleared land proportions (Fig. 3c, d) and flatter terrain (Fig. 3e, f) were significantly related to lower probability of loss and proportions of loss. Taller canopy height was significantly associated with greater probability of loss in both models (Fig. 3g, h). Full model test statistics and term significance are shown in Table 3. Comparing a structure with predictors indicative of high loss likelihood to one with low likelihood (Fig. 4a) yielded model estimates of 0.91 and 0.07, respectively. For proportional loss, estimates were of similar magnitude with 0.89 for a high-risk area and 0.16 for a lower-risk area (Fig. 4b).

Fig. 3.

Model estimated effects (holding other predictors at their means) are shown for structure count (top row, a and b), cleared land proportion (second row, c and d), slope (third row, e and f) and canopy height (fourth row, g and h). Binomial model estimates are shown at left and proportional loss estimates at right. In the binomial model plots, boxplots indicate the data distributions for losses (red shaded) and non-losses (blue shaded). In the proportional model plots, the observations are shown as points. Grey shaded areas on the model estimates are the 95% CI.

Fig. 4.

Comparison of low and high loss likelihood combinations of predictors variables for the binomial (a) and proportional model (b) as an illustrative example of the combined effects of the model predictors. The predictor values used are shown on the plots below or above the model estimates. Where there was a positive relationship between loss and a given predictor, the 95th percentile value from the dataset was used for the high-loss likelihood estimate and the 5th percentile for the low loss likelihood estimate. Where the relationship between loss and a predictor was negative this process was inverted.

Predictive accuracy for the binomial model was consistent between the fitted (Fig. 5a) and validation datasets (Fig. 5b), with both achieving a TPR of 0.67 and a TNR of 0.69. For both datasets, the median predictions showed reasonable separation between loss and non-loss cases, meaning the model was able to distinguish between the two on average. For the proportional model, r² was higher for the validation dataset (0.61 vs 0.71; Fig. 5d), compared to the modelled data (Fig. 5c), MAE was similar between datasets (0.14 fitted vs 0.12 validation) and bias remained unchanged (0.04).

Fig. 5.

Boxplot and density distributions of the binomial (Bin.) model predictions for the fitted data (a), and the validation data (b) disregarding information in the random effect grouping factor in both cases. In (a) and (b) counts of correctly predicted losses (red shaded) and non-losses (blue shaded) are shown with true positive rate (TPR) and true negative rate (TNR) proportions. Predictions for the proportional (Prop.) model on the fitted data are shown in (c), and for the validation data in (d). In (c) and (d), r², MAE and bias (prediction – observation).

The mapped model outputs show high median loss estimates (Fig. 6c; aggregated predictions for all structures in each grid cell) in areas with structures located within mapped fuel (Fig. 6a, b), highlighting the effect of cleared land on reducing the probability of structure loss. Maximum loss estimates were sensitive to structure count (Fig. 6d), returning lower estimates in areas with more structures. Proportional loss estimates (Fig. 6e; predictions made using cell median values for predictors) gave visually similar results to the median loss estimates (Fig. 6c).

Fig. 6.

Maps showing the Australian Fire Danger Rating System (AFDRS) fuel types (a), counts of structures (b), the median (c) and maximum (d) model loss estimates for these structures, and the proportional loss estimate model predictions (e). A key for (a) is shown above (a) with white corresponding to land outside the defined fuel types shown in the key, and colour scales for (b) to (e) shown to the right of each plot indicate values.

Overall, both the binomial and proportional models showed good performance in predicting structure loss at the individual level (binomial model) and the proportions of structures lost at larger spatial scales. Model predictions made on the AFDRS grid show clear spatial patterns in the estimates of structure loss, with the binomial model offering different methods to aggregate the predictions to grid scale (median and maximum probability; Fig. 6).

Discussion

The models described here are designed to predict the factors which may predispose structures to loss during fire. With this aim we focussed on the influence of landscape factors and structure counts as descriptors of structure loss. Both models retained terms for structure count (within 250 and 750 m for the binomial and proportional model respectively), cleared land within 400 and 100 m, canopy height within 400 and 50 m and TRI. Taken together, these indicate lower probability of structure loss where land (a) is 100% cleared surrounding the structure in a buffer (defensible space), (b) has shorter canopy height in surrounding vegetation, (c) is on flat ground, and (d) has a higher number of structures nearby. These results are in concordance with other studies which found defensible space (Syphard et al. 2014), TRI (Price and Bradstock 2013) and surrounding structure density (Syphard et al. 2012) which are important predictors of structure loss. Notably, this study found a larger defensible space was required (buffer distance between the structure and adjacent fuel) – 400 m for the binomial model and 100 m for the proportional model – compared to previous studies. In contrast, Gibbons et al. (2012) and Syphard et al. (2014) reported 40 m, and up to 30 m, respectively. This difference is likely due to the use of relatively low-resolution spatial data used in this study to characterise surrounding land compared to the previous research. Studies which directly measure these properties with respect to the direction of fire attack will give a more accurate estimate for the buffer sizes of defensible space. In this study, buffer sizes are likely influenced by data resolution, capturing vegetation surrounding a structure rather than only in the direction the fire travelled before causing its loss. The 1-km proportion of forested land identified by Price and Bradstock (2013) as a predictor of house loss, was calculated similarly to the cleared land in this study, suggesting that this difference in methodology may account for the variation between studies.

The binomial model showed reasonable classification accuracy (TPR = 0.67 and TNR = 0.69 on the validation data; Fig. 5). The proportional loss model also showed a good predictive fit (r² = 0.71 on the validation data and 0.61 on the calibration/fitting data).

Other regression model-based studies have identified different primary contributors to their predictions. For example, Wilson and Ferguson (1986) and Blanchi et al. (2010) used the Forest Fire Danger Index (FFDI) to infer house loss, while Gibbons et al. (2012) focussed on fuels and their attributes (cover within 40 m of houses, fuel type, upwind distance to fuels and fuel treatment). Harris et al. (2012) found that adjusted fire danger indices and fire relative energy were the best predictors of house loss, and Duff et al. (2019) incorporated convection, ember density and flame height as the sole contributors to their regression-based house loss model, while Duff and Penman (2021) added perceived slope, presence of fuels within 40 m and percentage of fuel within 100 m as predictors of loss.

Predictors of house loss can be broadly classified into weather-, fire behaviour-, or fuels-based factors. In the model presented here, we used fuels-based, topographic (TRI) risk factors, and structure counts to predict structure loss. This approach has allowed the development of a model which aligns well with the AFDRS operational framework and can be interpreted alongside AFDRS forecasts, e.g. the FBI. To ensure compatibility with existing AFDRS models and data, the impact forecasts generated in this study are aggregated to the AFDRS 1.5 × 1.5 km grid (Fig. 6).

Conclusion

The impact index aims to provide decision makers with an objective likelihood forecast that will provide output in a format consistent with other AFDRS products (e.g. FBI), with which it can be combined into a fire behaviour weighted output and is designed for use by decision makers and Fire Behaviour Analysts. In an Incident Management Team, objective metrics provide valuable intelligence that can be used to temper more subjective assessments of what is likely to happen during a running fire. While an index can only provide an estimate (with inherent uncertainty), the introduction of the impact index will make a significant contribution to tools available to those making important decisions regarding wildfire control in Australia. This model is a first step in providing this information and will be refined or replaced as understanding of the drivers of structure loss during wildfire progresses.

Supplementary material

Supplementary material is available online.