Since we are investigating several different indicator or driver variables of fire risk, different event definitions are developed for different variables. The details of those definitions are given at the beginning of the respective sections on temperature, precipitation and fire weather indices (Sects. –). General parameters of the event definition are given here.

The fire season (September–February) serves as the general event time window, and the region with the most intense fires in 2019/20 in southeastern Australia serves as the general event spatial domain; specifically this is the land area in the polygon 29∘ S, 155∘ E; 29∘ S, 150∘ E; 40∘ S, 144∘ E; and 40∘ S, 155∘ E (as shown in Fig. ), which corresponds to the area between the Great Dividing Range and the coast.

The primary way we investigate the connection between anthropogenic climate change and the likelihood and intensity of dangerous bushfire conditions is through the FWI. The FWI provides a reasonable proxy for the burned area in the extended summer months, with the strongest relationship observed from November to February. Figure  shows both the Spearman rank correlation and the Pearson correlation of the FWI with log-transformed burned area. The 95 % confidence intervals are also shown. Given the similarity in the correlation coefficients (r) within their confidence intervals, the log-linear relationship appears to explain equal variability (r2) to that of the ranks.

To capture spatial variations in the start of the fire season at a given location within the event domain, we take for most quantities first the maximum per grid point over the fire season (September–February) and next the spatial average over the general event domain. This way the events do not need to be simultaneous at separate grid points within the region. We therefore investigate the question how anthropogenic climate change influences the chances of an intense bushfire season, rather than focusing on a single episode of intense bushfires.

In most years only very small areas are burned, but the observational record also includes events with extremely large areas. Given this, we checked if the burned-area observations were heavy-tailed . We found that monthly burned area was not Pareto distributed and instead is reasonably approximated using a log-normal distribution. This supports using the log transformation and extrapolating this relationship to the 2019/20 fire season. Temporal detrending of the observations did not alter these conclusions.

The observational data used in this study are described in Sects. S1 and S2 in the Supplement and  for heat, drought and the Fire Weather Index, respectively, including justifications for including or excluding certain datasets for certain research questions. For the global mean surface temperature (GMST) we use GISTEMP (Goddard Institute for Space Studies Surface Temperature Analysis) surface temperature .

Attributing observed trends to anthropogenic climate change can only be done with physical climate models, as they allow for isolating different drivers. For this purpose we included as large a set of ocean–atmosphere coupled and atmosphere-only (i.e. sea surface temperature (SST) prescribed) climate model ensembles as we could find within the time constraints of this study in order to obtain estimates of both the uncertainty due to natural variability and the model uncertainty. A selection of large ensembles of climate models from the Coupled Model Intercomparison Project Phase 5 (CMIP5) has been used: CanESM2 (Canadian Earth System Model), CESM1-CAM5 (Community Earth System Model–Community Atmosphere Model), CSIRO Mk3.6.0 (Commonwealth Scientific and Industrial Research Organisation), EC-Earth 2.3 (European community earth system model), GFDL CM3 (Geophysical Fluid Dynamics Laboratory Climate Model), GFDL ESM2M (GFDL Earth System Model 2) and MPI-ESM (Max Planck Institute for Meteorology Earth System Model). In addition, the HadGem3-A N216 (Hadley Centre Global Environment Model) attribution model developed in the EUropean CLimate and weather Events: Interpretation and Attribution (EUCLEIA) project, the weather@home (HadAM3P, Hadley Centre atmosphere model) distributed attribution project model, and the ASF-20C (Atmospheric Seasonal Forecasts of the 20th Century) seasonal hindcast ensemble have been used. These last three models are uncoupled and forced with observed historical SSTs and estimates of SSTs, as they might have been in a counterfactual world without anthropogenic climate change. Finally, we used the coupled IPSL-CM6A-LR (Institut Pierre‐Simon Laplace Climate Model) low-resolution CMIP6 ensemble. The GFDL-CM3 and MPI-ESM models that did not have daily data were not used for the extreme-heat analysis. A list of these climate models and their properties is given in Table 1. For the FWI analysis, which requires daily data of relative humidity (RH), temperature, precipitation and wind speed, the list of models used is shortened to CanESM2, CESM1-CAM5, EC-Earth, IPSL-CM6A-LR and weather@home.

The methods employed in this analysis have been used previously for high and low temperatures , extreme precipitation , drought and forest fire weather . A paper describing the methods in detail was recently published as .

Changes in the frequency of extreme events are calculated by fitting the data to a statistical distribution. In this study the highest temperature extremes and fire-risk-related variables (FWI and MSR) of the fire season are assumed to follow a generalized extreme value (GEV) distribution, which is the distribution that block maxima converge to . While our event definition is not exactly block maxima, the GEV fits the data well (see below for more details). The low values of annual mean precipitation and lowest monthly precipitation of the fire season are fitted using a generalized Pareto distribution (GPD), which describes the exceedance below a low threshold and also allows for the specification of a threshold that ensures the PDF (probability density function) is zero for negative precipitation.

The GEV distribution is 1P(x)=exp⁡-1+ξx-μσ-1/ξ, where x the variable of interest, e.g. temperature or precipitation. Here, μ is the location parameter; σ>0 is the scale parameter; and ξ is the shape parameter. The shape parameter determines the tail behaviour: a negative shape parameter gives an upper bound to the distribution, for ξ≥1 the tall is so fat that the mean is infinite. The scale parameter corresponds to the variability in the tail.

The GPD gives a two-parameter description of the tail of the distribution above a threshold, where the low tail of precipitation is first converted to a high tail by multiplying the variable by -1. The GPD is then described by 2H(u-x)=1-1-ξxσ(-1/ξ), with x being the temperature or precipitation, u being the threshold, σ being the scale parameter, and ξ being the shape parameter determining the tail behaviour. For the low extremes of precipitation, the fit is constrained to have zero probability below zero precipitation (ξ<0, σ<uξ). Calculations were conducted on the lowest 20 % and 30 % of the data, which provide a first-order estimate of the influence of using more or less extreme events. We cannot use less data, as the maximizations of the likelihood function do not converge anymore, and using more than 30 % would not qualify as the “lower tail”.

Drought (or low precipitation) is particularly difficult to model using the existing extreme value framework . While minima can be modelled by multiplying by -1 , the applicability of the underlying extreme value theory assumptions still needs to be validated. In the case of low precipitation, year-on-year autocorrelations are a concern. In southeastern Australia, these serial autocorrelations are approximately r≈0.2, so although they are non-zero, they do not dominate the drought characteristics. Despite these theoretical limitations, in practice the diagnostic plots show that the generalized Pareto models are able to describe the data reasonably well. In particular, they respect that precipitation is non-negative. In general this is a difficult problem, and the statistical extremes community is developing solutions necessary for modelling drought events .

To calculate a trend in transient data, some parameters in these statistical models are made a function of the 4-year smoothed global mean surface temperature (GMST) anomaly T′. This smoothing is the shortest that on the one hand reduces the ENSO component of GMST, which is not externally forced and therefore not relevant for the trend, but on the other hand it retains as much of the forced variability as possible . A longer smoothing timescale would create problems with extrapolation in the highly relevant last few years of the instrumental record. The covariate-dependent function can be inverted and the distribution evaluated for a given year, e.g. a year in the past (with T′=T0′) or the current year (T′=T1′). This provides estimates and confidence intervals of the probabilities for an event at least as extreme as the observed one in these 2 years, p0 and p1, or expressed as return periods τ0=1/p0 and τ1=1/p1. The change in probability between 2 such years is called the probability ratio (PR): PR=p1/p0=τ0/τ1. We also estimate the changes in intensity (including uncertainties): ΔT for temperature, ΔP for drought and ΔFWI.

For extreme temperature we assume that the distribution shifts with GMST as μ=μ0+αT′ or u=u0+αT′ and σ=σ0 with α denoting the trend, which is fitted together with μ0 and σ0. The shape parameter ξ is assumed constant. For drought and FWI-related variables we instead make the assumption that the distribution scales with GMST, the scaling approximation . In a GEV fit this gives 3μ=μ0exp⁡αT′/μ0,σ=σ0exp⁡αT′/μ0, and in a GPD fit, it is u=u0exp⁡αT′/u0,σ=σ0exp⁡αT′/μ0, with fit parameters σ0, α and ξ. The threshold u0 is determined with an iterative procedure, and the shape parameter ξ is again assumed constant. The exponential dependence on the covariate is in this case just a convenient way to ensure a distribution that is zero for negative precipitation and has no theoretical justification. For the small trends in this analysis it is similar to a linear dependence.

The validity of the other assumption, that the scale parameter or dispersion parameter are constant, is tested by computation of the significance of deviation of a constant of running (relative) variability plots of the observations and model data . The analysis of model data is more sensitive to variations of these parameters over time due to the large number of ensemble members but of course assumes the effect of external forcing on the variability is modelled correctly.

For all fits we also estimate 95 % uncertainty ranges using a non-parametric bootstrap procedure, in which 1000 derived time series, generated from the original one by selecting random data points with replacement, are analysed in exactly the same way. The 2.5th and 97.5th percentile of the 1000 output parameters (defined as 100i/1001 with i being the rank) are taken as the 95 % uncertainty range. For some models with prescribed SSTs or initial conditions (in the case of the seasonal forecast ensemble) the ensemble members are found to not be statistically independent, defined here by a correlation coefficient r>1/e with e≈2.7182. In those cases the same procedure is followed except that all dependent time series are entered together in the bootstrapped sample, analogous to the method recommended in to account for temporal dependencies.

When using a GEV to model tail behaviour, note that taking the spatial average of the annual maxima does not have the same statistical justification as taking the annual maximum of the spatial average . Given this, the impact of the order of operations in the event definition was examined. For the temperature extremes, we compared the time series where we first take the annual maximum and next the spatial average to the definition with the order reversed, which can be approximated with a GEV. The Pearson correlation was r=0.95, which is likely due to strong spatial dependence and the concentration of heatwaves at the peak of the seasonal cycle. Therefore, in practice, an approximation with a GEV is not entirely unsuitable for temperature, but caution should be exercised. For the FWI and MSR, the order of operations does make a clear difference. Indeed, we find that the whole distribution is not described well by a GEV for one climate model used (CanESM2). For that model we take block maxima over five ensemble member blocks, effectively looking only at the most extreme events. For this part of the distribution the GEV fit agrees with the data points in the return time plot, as expected from taking block maxima.

We evaluate all climate models on the fitting parameters by determining whether the model-derived parameters fall within the uncertainty range of observation-derived parameters. We allow for a mean bias correction; i.e. we only check the scale and shape parameters σ and ξ. Model biases are accounted for by evaluating the model at the same return time as the value found in the observational analysis. This was found to give better results than applying an additive or multiplicative bias correction to the position parameter μ, as it also corrects to first order for biases in the other parameters, especially when the distribution has an upper or lower bound (ξ<0), which is the case in all the cases here.

Finally, estimates of the PR and change in observations and all climate models that pass the evaluation test are combined to give a synthesized attribution statement. First, the observations and reanalyses were combined by averaging the best estimate and lower and upper bounds, as the natural variability is strongly correlated, as they are largely based on the same observations (except for the long reanalyses). The difference is added as representation uncertainty (white extensions on light-blue bars in Figs. S6, S12, S13 and ).

Second, the model results were combined by computing a weighted average (using inverse model total variances), as the natural variability in the models, in contrast to the observations, is uncorrelated: 4X‾=∑iXi/σXi2/∑1/σXi2, with σXi being the estimated uncertainty in model i and Xi being either the temperature or the logarithm of precipitation, FWI or MSR. The sums are over the Nmod models. Using this we can compare the spread expected from the natural variability with the observed spread of the model results using χ2 statistics: 5χ2=∑iXi-X‾i/σXi. If χ2/dof≤1, with dof being the number of degrees of freedom, here N-1, the spread of the results is compatible with the uncertainty estimated from the fits due to variability within the climate model, and the results can be taken to be independent estimates of X and the weighted average used. However, if χ2/dof>1 the model spread is larger than expected from variability due to sampling of weather noise alone, so a model spread term was added to each model in addition to the weighted average (white extensions on the light-red bars, Figs. S5 and S6) to account for systematic model errors. This term is defined by requiring that χ2/dof=1. The total uncertainty of the models is shown as a bright-red bar in these figures. This total uncertainty consists of a weighted mean using the uncorrelated natural variability plus an independent model spread term added to the uncertainty if χ2/dof>1, which we do not divide by N-1; i.e. we do not assume that by adding more models to the ensemble the model uncertainty decreases. This procedure is similar to the one employed by .

Finally, observations and models are synthesized into a single mean and uncertainty range. This can only be done when they appear to be compatible. We show two combinations. The first one is computed by neglecting model uncertainties beyond the model spread. The optimal combination is then the weighted average of models and observations, shown as a magenta bar. However, the total model uncertainty is unknown and can be larger than the model spread. We therefore also show the more conservative estimate of an unweighted average of observations and models with a white box in the synthesis plots.

As discussed in the introduction, the fire risk as described by fire weather indices was extreme in the study domain in the 2019/20 fire season. The domain was chosen to encompass these fires, and therefore the 2019/20 event cannot be included in the statistical analysis.

We choose two event definitions in order to represent two important aspects of the event, namely the intensity and the duration. For intensity, we first select the maximum FWI of a 7 d moving average over the fire season (September–February) for every grid point over the study region, after which we compute the spatial average, hereafter FWI7x-SM (seasonal maximum). The 7 d timescale was chosen based on a good correlation with the area burned (see Fig. ) and good correspondence with area burned in other forest fire attributions studies .

For duration, we consider the monthly severity rating (MSR). The MSR is the monthly averaged value of the daily severity rating (DSR), which in turn is a transformation of the FWI (DSR=0.0272FWI1.71). The DSR reflects better how difficult a fire is to suppress, while the MSR is a common metric for assessing fire weather on monthly timescales . For this study, we select the maximum value of the MSR during the fire season over the study area (MSR-SM). In contrast to the FWI7x-SM, we first apply a spatial average of the study area and then select the maximum value per fire season. This event definition focuses more on changes in extreme fire weather for longer timescales and larger integrated areas than FWI7x-SM. Note that neither of the two event definitions includes ignition sources or small-scale meteorological factors such as pyrocumulonimbus development that could enhance the fires.

For the observational analysis we use the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) Atmospheric Reanalysis (ERA5) dataset for 1979–January 2020 . This reanalysis dataset is heavily constrained by observations and thus provides one of the best estimates of the actual state of the atmosphere for all the variables needed to compute the FWI over the study area. Other reanalyses did not yet include the full 2019/20 event at the time of the analysis.

Figure  shows the time series of the highest 7 d mean FWI averaged over the study area. Both for the FWI7x-SM and MSR-SM the event is the highest over the 1979–2020 time period. Note that for the MSR-SM, the value is considerably more extreme than for the FWI7x-SM. The GEV fits (Fig. , right) illustrate this further, with return times in excess of 1000 years.

A fit allowing for scaling with the smoothed GMST gives a significant trend in the FWI7x-SM (Fig. ). This fit gives a return time for the 2019/20 fire season of about 31 years (4 to 500 year) in the current climate and more than 800 years extrapolated to the climate of 1900. This corresponds to an infinite PR, with a lower bound of 4. The return time for the MSR-SM is undefined and is thus estimated to be 100 years. For the climate model analysis we thus use return times of 31 years for the FWI7x-SM and 100 years for the MSR-SM to determine the event thresholds in individual climate models.

We use four climate models with large ensembles, leaving out CESM1-CAM5 because of its failure to represent heat extremes (see Sect. ). This is fewer than for the drought and heat analysis because the FWI requires four daily input variables, which are not available for all models. In contrast to the heat extremes and drought analyses, the fits to the model output use as covariate the model GMST. We also define our reference climates using GMST rather than years. The years at which the climate is evaluated are taken from the 1.1 ∘C temperature increase for the present-day climate and the 2 ∘C increase for the future reference climate and not 2019 and 2060. As the fits are invariant under a scaling of the covariate, this does not make much difference.

First the models are evaluated on how well they represent the extremes of the FWI7x-SM and MSR-SM. This is quantified by the dispersion parameter σ/μ and shape parameter ξ of the GEV fit for the present-day climate. We do not check the position parameter μ, assuming a multiplicative bias correction can be applied.

Figure  gives an overview of these parameters. Preferably, we would like the parameters to lie within the observational uncertainty of ERA5. For the dispersion parameter CanESM2 and weather@home fall within the observational uncertainty of the FWI. The other two models (EC-Earth and IPSL CM6) show too much variability relative to the mean. The same holds for the shape parameter. This implies that it is difficult to draw strong conclusions from the model data, given that they do not accurately represent the extremes of the FWI7x-SM. In particular, the models with too much variability will underestimate the probability ratios. We continue with all four models but keep these problems in mind.

The MSR is simulated better: all model dispersion and shape parameters lie within the large observational uncertainties, although they largely disagree with one another on the dispersion parameter.

The model results are summarized by their PR, i.e. how more or less likely such an event will be for present or future climate, relative to the early 20th century.

Figures  and  show the change in probability for both the FWI7x-SM and the MSR-SM from 1920 to 2019 (denoting the 2019/20 fire season). For the FWI7x-SM, all models agree on an increased probability for such an event in the present climate relative to the early 20th century, although the trend is not significant at p<0.05 (two-sided) for one of the models, CanESM2. As the spread of the models is compatible with natural variability (χ2/dof<1), we take a weighted average across the models to synthesize them (Fig. ). This shows that such an event has become about 80 % more likely in the models, with a lower bound of 30 %. Note that all models severely underestimate the increased risk compared to ERA5, which has a lower bound of the PR of a factor of 4 relative to 1920 (extrapolated), above the upper end of the model average. Note that the ERA5 value is probably biased high, as the positive contribution of trend towards a drier climate over 1979–2019 is not present over 1900–2019; see Sects. S2 and .

For a future climate of 2 ∘C warming above pre-industrial levels we find that such events become about 8 times more likely in the models, with a lower bound of about 4 times more likely. Note that the estimate of future climate is only based on two climate models, CanESM2 and EC-Earth, due to the absence of future data for the others.

For the MSR-SM the models on average show about a doubling of probability for the present climate relative to the early 20th century (Fig. ). However, this trend is not significant, as the lower bound is 0.8; i.e. a decreased probability is also possible within the two-sided 95 % uncertainty range. In the fit to the ERA5 data we include 2019, as otherwise the probability of the event occurring in the current climate would be zero, contrary to the fact that it did occur. This fit shows much higher probability ratios, with a lower bound of a factor of 9. As there is no overlap with the model results we cannot combine the model and observational results but only give a conservative lower bound as an observation–model synthesis result. For a future climate relative to the climate of the early 20th century the models show an increase in probability of about 4 times, with a lower bound of 2.

The underestimation of the observed trend in fire weather indices in all models and the tendency for too much variability in some models is reminiscent of the extreme-temperature results in Sects.  and S1.

In order to better understand which input variables cause the long-term increase in the FWI7x-SM and thus the contribution to the 2019/20 FWI7x-SM value, we study the input variables to the FWI7x-SM separately for each model as well as observations. For precipitation we use the cumulative precipitation (90 d) prior to each FWI7x-SM value. We calculate the change from the early 20th century to the present day in each input variable to estimate its long-term change, which we then subtract from that variable's observed value in 2019/20. We then recalculate the FWI7x-SM but use each detrended individual input variables in turn. Each of these newly calculated FWI7x-SM values thus illustrates the influence of the long-term trend in a particular input variable onto the observed 2019/20 FWI7x-SM value. This procedure is applied to models and ERA5. In the models, the ensemble mean change is used to estimate an individual variable's long-term trend, whereas in ERA5 a regression of each variable onto GMST is used to estimate its value in the early 20th century. The results of this analysis for the 2019/20 FWI7x-SM value are shown in Fig. .

The sum of the contributions from individual input variables to the 2019/20 FWI7x-SM anomaly match the effect of changing all variables at the same time, so they can be considered linearly additive (Fig. ). The underestimation of the extreme-temperature trends in the climate models carries over into this analysis such that the temperature contribution to the observed 2019/20 value is underestimated. Despite this underestimation, temperature emerges as the most important variable in EC-Earth and weather@home, as it explains roughly half of the increase in the FWI. For IPSL, the simulated temperature increase explains about a third of the FWI7x-SM increase, together with wind and RH. CanESM2 behaves differently, where it is mainly the decrease of RH that explains the higher FWI7x-SM. Most but not all models analysed here therefore derive the increase in the FWI7x-SM largely from the increase in temperature extremes.

In ERA5 the increase in temperature also appears to be the most important explanatory variable, followed by a decrease in RH and precipitation. As we did not find a significant trend in precipitation over longer time periods, we hypothesize this trend to be due to natural variability over the short 1979–2018 period in ERA5. We explicitly verified that the dependence of the FWI7x-SM on temperature is almost linear in a range of ±5 K around the reanalysis value (not shown). Further, volumetric soil water (Fig. ) at multiple soil layers from ERA5 suggests that, despite the soil already being very dry in 2018 and into 2019, the 2019/20 austral spring–summer drought caused a further drying of the soil in the study area. This suggests that the drought of late 2019 and high temperatures did indeed cause an additional increase in fire risk over preceding years.

The FWI7x-SM as computed from the ERA5 reanalysis as an approximation to the real world shows that the 2019/20 values were exceptional. They have a significant trend towards higher fire weather risk since 1979. Compared with the climate of 1900, the probability of an FWI7x-SM as high as in 2019/20 has increased by more than a factor of 4. For the MSR-SM the probability has increased by more than a factor of 9.

The four climate models investigated show that the probability of a Fire Weather Index this high has increased by at least 30 % since 1900 as a result of anthropogenic climate change. As the trend in extreme temperature is a driving factor behind this increase and the climate models underestimate the observed trend in extreme temperature, the attributable increase in fire risk could be much higher. This is also reflected by a larger trend in the FWI7x-SM in the reanalysis compared to models. The MSR-SM increased by a factor of 2 in the models since 1900, although this increase is not significantly different from zero. As with FWI7x-SM, the attributable increase is likely higher due to the model underestimation of temperature trends and overestimation of variability in the TX7x.

Projected into the future, the models project that an FWI7x-SM as high as in 2019/20 would become at least 4 times more likely with a 2 ∘C temperature rise, compared with 1900. Due to the model limitations described above this could also be an underestimate.

The time of publication is too soon for a robust estimate of excess morbidity and mortality specific to the 2019/20 Australian bushfires. Such analysis is typically available weeks to years following the end of an event. However, the combined impacts of extreme heat and air pollution can be deadly, as seen in the compounded heatwave and wildfire events in 2010 in Russia or 2015 in Indonesia (; https://www.nature.com/articles/news.2010.404, last access: 7 March 2021; ). Those most at risk are the elderly; people with pre-existing cardiovascular, pulmonary and/or renal conditions; and young children, as exposure to wildfire smoke can have acute respiratory effects. Health officials in New South Wales reported a 34 % spike in emergency room visits for asthma and breathing problems between 30 December 2019 and 5 January 2020 (https://www.washingtonpost.com/climate-environment/2020/01/12/australia-air-poses-threat-people-are-rushing-hospitals-, last access: 7 March 2021). One study of hospital admissions in Sydney, Australia, from 1994 to 2010 found that days with air pollution from extreme bushfires (as measured by PM10) resulted in a 1.24 % admission increase for every 10 µg m-3 . On the other hand, it should be noted that Australia is a country with a robust healthcare system, which significantly reduces vulnerability to the short- and long-term consequences of smoke and extreme heat.

There is also a need for increased mental health services in the days, weeks and years following severe bushfires. As of January 2020 the Australian government announced AUD 76 million in mental health funding (https://www.abc.net.au/news/2020-01-12/federal-government-funds-for-mental-health-in-fires-crisis/11860660, last access: 7 March 2021). A study of the 2009 Black Saturday bushfires found that while the majority of affected people demonstrated psychological resilience in the long-term aftermath of the fires, a significant minority of people in highly affected communities reported mental health impacts 3–4 years following the event .

There is no nationally standardized system for bushfire warnings in Australia. However, recommendations from the royal commission tasked with reviewing the 2009 Victoria bushfire have helped to drive forward efforts to establish a national system. In 2014 the National Review of Warnings and Information was undertaken. It recommended the establishment of the dedicated, multi-hazard National Working Group for Public Information and Warnings. Part of the task of this group would be to ensure greater national consistency of early warning information. One outcome of this recommendation is the Public Information and Warnings Handbook, which has been issued to provide guidance to actors across national, state and territory governments in issuing warning information (https://knowledge.aidr.org.au/media/5972/warnings-handbook.pdf, last access: 7 March 2021).

Bushfire warnings in Australia are issued by state and territory fire authorities and generally follow the “Prepare, Stay and Defend or Leave Early” approach. The Fire Danger Rating system is also widely applied as a way to communicate fire risks. The system, originally developed in 1967, contained five risk levels ranging from “low-moderate”, where fires can generally be controlled, to “extreme”, where evacuation is recommended but home defence may be possible under certain circumstances. Following the 2009 Black Saturday bushfires a sixth “catastrophic” level was added where evacuation is deemed the only survival option (https://www.emergency.wa.gov.au, last access: 7 March 2021). This guidance was adopted by all states, except in Victoria where it is called “Code Red” (http://www.bom.gov.au/weather-services/fire-weather-centre/fire-weather-services/index.shtml, last access: 7 March 2021). In 2017, the system was revised to update the metrics used in forecasting the most appropriate level (https://www.abc.net.au/news/2017-12-13/bushfire-danger-rating-system-trialled-summer/9203446, last access: 7 March 2021). Threat level information is provided via radio, television, social media and signs on all major rural roads. Government websites also provide information which is updated every few minutes and includes maps of fires and associated threat levels. In addition, phone calls are made house-to-house when evacuation is recommended.

While these efforts help to reduce vulnerability and exposure to the wildfires, significant barriers to early action still remain. People in bushfire areas are frequently not aware of their risk, are unprepared to manage risk, wait until the final moments to evacuate or, at times, even return to fire-affected areas to defend property . This is particularly relevant in peri-urban areas which are not as frequently exposed to bushfire risks. A 2020 study of people's reactions to bushfire warnings during the 2017 bushfires in New South Wales found that people largely understood warnings but that they did not respond to the warnings before seeking and obtaining additional confirmation to avoid what they might perceive as unnecessary evacuation and associated costs. Researchers recommend that rather than further refining messages, confirmation mechanisms need to be implemented into early warning approaches. There are further barriers to acting on warnings, such as an incorrect assessment of the defensibility of a building, attachments to pets and personal items, and a “hero culture” around people who did defend their home. These sociological barriers to life-saving measures increase the risk of deadly impacts.

Furthermore, an inquiry into the 2009 Black Saturday bushfires found that the Prepare, Stay and Defend or Leave Early approach assumed that individuals had a fire plan in-place, although many people did not. Therefore people were left in a position to make complex decisions without adequate guidance.

For this study, we did not assess vegetation cover and condition (dryness) ahead of the season in comparison with earlier years, but it is clear that fire hazard management strategies such as bush reduction, manual removal of undergrowth and controlled burning can affect fire hazard

Note that the effectiveness of such measures depends not only on the type of vegetation but also the specific conditions of the bushfire. For instance, prior controlled burning may be somewhat effective to suppress fire risk under average weather conditions but is much less so in cases of very high temperatures, low humidity and strong wind, i.e. when the fire risk is no longer dominated by the type, condition and quantity of fuel but by weather conditions. During the 2009 fires in Victoria, recently burned areas (up to 5–10 years) may have reduced the intensity of the fires, but they did not enough so as to increase the chance of effective suppression given the severe weather conditions at the time .

In addition, it should be noted that controlled burning requires a window during the cooler parts of the year when conditions allow for controlled burning to take place. The Queensland Fire and Emergency Services (QFES) noted that controlled burning is highly dependent on weather conditions and that not all planned 2019 burns had been completed, given that in some areas, it rapidly became too dry to burn safely (https://www.abc.net.au/news/2019-12-20/hazard-reduction-burns-bushfires/11817336, last access: 7 March 2021). Recognizing the highly non-linear relationship between weather conditions through the season (and in fact across several years) and anticipatory risk management strategies, in this attribution study we have not assessed the impact of these early-season weather conditions on the ability to reduce risk and thus on fire risk itself.

Ageing electricity infrastructure may play a role in increasing the risk of bushfire outbreaks from human-related ignition . Electric-grid fires are primarily due to elastic extension and fatigue failures and are made increasingly worse by high wind speeds . A 2017 study found that fires sparked by electricity failures are more prevalent during elevated fire risk and tend to burn larger, making them worse than fires due to other causes . Interestingly, a 2013 report also notes that while electricity operators have the ability to disconnect electricity grids when there is a high risk posed to the public, only South Australia has legislation in place to protect the operator from prosecution . All of these factors coupled together increase Australia's vulnerability to bushfire outbreaks.

In contrast, stringent building codes have helped to reduce vulnerability to fire risks. The Bushfire Attack Level (BAL), as well as associated building codes, is the guiding resource for assessing and managing risks of a building exposed to heat, embers or direct fire. The BAL is applied nationwide; however the Fire Danger Index, one of the key metrics used in calculating the site-specific BAL, is under state and local level jurisdiction. The highest BAL level was established following the 2009 Black Saturday bushfires .

Land use planning at a community level is also crucial in reducing bushfire risk, particularly for rural and peri-urban areas which face the highest bushfire risks. This is recognized and addressed through state and local government planning processes, which include ensuring accessible bushfire evacuation routes and spaces. For example, the 2009 Victorian Bushfires Royal Commission cited a need for planning which “prioritized human life over all other policy objectives”. This led to relevant policy changes through an amendment to the Victoria Planning Provisions. The Bushfire Management Overlay and associated guidelines are among the principle aspects of this amendment. They provide direction for approval of new construction locations as well as siting and layout requirements of approved spaces, although these guidelines do not apply to existing property, which puts a limitation on their overall positive impact .

Bushfires are a natural phenomenon, but their impact is also strongly influenced by human choices. Overall, Australia is one of the most prepared countries in the world to manage bushfires, and thus the impacts from this season's bushfire outbreaks could have been dramatically worse if not for the systems in place. This not only underscores the urgent need to adapt to changing risks in all places, with a special focus on the most vulnerable, but also highlights the limits to risk reduction and preparedness.

As a result of the Black Summer bushfires, formal inquiries have been launched in Victoria (https://www.igem.vic.gov.au/vicfires-inquiry, last access: 7 March 2021), New South Wales (https://www.nsw.gov.au/nsw-government/projects-and-initiatives/nsw-bushfire-inquiry, last access: 7 March 2021), Queensland (https://www.igem.qld.gov.au/queensland-bushfires-review-2019-20, last access: 7 March 2021) and South Australia (https://www.safecom.sa.gov.au/independent-review-sa-201920-bushfires/, last access: 7 March 2021). A federal royal commission has been announced with an aim to improve resilience, preparedness and response to disasters across all levels of government. The commission will also seek to improve disaster management coordination across local government and improve relevant legal frameworks. These inquiries will shed additional light on the vulnerability and exposure elements of the Black Summer bushfires and hopefully help mitigate future risk.