What environmental and climatic factors influence multidecadal fire frequency?

Abstract

Fire is a key ecosystem process with more than half the world's land surface potentially subject to fire. A key aspect of fire ecology is the fire regime, with fire frequency an important component. Fire frequency appears to be increasing in some ecosystems, but decreasing in others. Such temporal and spatial variability in fire frequency highlights the importance of more effectively quantifying spatiotemporal changes in fire frequency for particular environments. We modeled changes in fire frequency over the past 40 years (1981–2020) in a 4.64 million ha area in Victoria, Australia. We quantified regional variation in the number of fires (hereafter termed fire frequency) during two 20-year time periods (1981–2000 vs. 2001–2020), employing the Interim Biogeographic Regionalisation for Australia (IBRA), a standardized regionalization of Australia's terrestrial landscapes. We also quantified the climate and environmental factors influencing fire frequency in each IBRA subregion. Our empirical analyses revealed that fire frequency in Victoria was heterogeneous in both time and space. Wildfire frequency changed between 1981 and 2020, with the past 20 years (2001–2020) experiencing a substantially greater number of fires relative to the 20 years prior (1981–2000). Changes in fire frequency were not spatially uniform, with increases more pronounced in some IBRA subregions than others. Climate and topographic factors influenced the frequency of wildfires, but their effects manifested differently in different IBRA subregions. For example, fire frequency was associated with increasing rainfall deficit deviation in four IBRA subregions, but an opposite trend characterized two others. Associations between fire frequency and increasing temperature deviation also varied from negative to positive across subregions. We also found evidence of elevation, slope, and aspect effects, but these too varied between IBRA subregions. The complex spatiotemporal changes in fire frequency quantified in this study, and the complex between-region differences in the factors associated with the number of fires, have major implications for biodiversity conservation, resource availability (e.g., timber yields), and ecosystem integrity. In ecosystems subjected to repeated fires at short intervals, new rapid detection and swift suppression technologies may be required to reduce the risks of ecosystem collapse as high-severity wildfires increase in frequency.

INTRODUCTION

Fire is a critical ecological process in many ecosystems globally (Bowman et al., 2009; Keeley et al., 2012; Keeley & Pausas, 2019). A fire regime, defined as the attributes of a sequence of fires in a given area (e.g., timing, frequency, severity, heterogeneity, and seasonality), is a key component of fire ecology (Gill, 1975; Keeley, 2009). Typically, a given ecosystem has a distinctive natural fire regime (Baker & Williams, 2018; McWethy et al., 2013) with which organisms have co-evolved (Whelan, 1995). In some forest types, such as Gondwanic rainforest and mangrove forests in eastern Australia, a natural fire regime can include a complete absence of fire (Kooyman et al., 2020; Lindenmayer, Bowd, & Gibbons, 2022). In other forest ecosystems such as Ponderosa pine (Pinus ponderosa) forests in northwestern North America (Larson et al., 2013), the natural fire regime consists of low-severity fires that occur every 5–30 years (Noel et al., 1998; Nowacki & Abrams, 2008; Stephens et al., 2016). In others, the fire regime may include rare, high-severity stand-replacing fires that would naturally occur every 75–150 years (such as in the Mountain Ash Eucalyptus regnans and Alpine Ash Eucalyptus delegatensis forests of southeastern Australia; McCarthy et al., 1999; Smith et al., 2016).

One of the key drivers of fire regimes and the behavior of individual fires is climate and weather, such as prolonged drought, high temperatures, limited humidity, and high windspeeds (Anderegg et al., 2022; Sullivan et al., 2012; Zylstra et al., 2022). Hence, there are strong links between the increased incidence of fires and climate change (Jones et al., 2020; Stephens et al., 2020). For example, long-term warming and increasingly dry climatic conditions in parts of North America and southeastern Australia appear to be having marked effects on fire regimes, which include increases in the frequency with which fires are occurring (Anderegg et al., 2022; Canadell et al., 2021; Singleton et al., 2019; van Oldenborgh et al., 2021).

While wildfires are becoming more frequent in many environments globally (Cattau et al., 2020; Collins et al., 2021; Jones et al., 2022; Levine et al., 2022; Mahood & Balch, 2019; Stewart et al., 2021; Westerling et al., 2006), in others their frequency has remained unchanged (McWethy et al., 2018) or even reduced, for example, as a result of rapid fire suppression (Palmquist et al., 2014; Zackrisson, 1977), or because of agricultural expansion and intensification (Andela et al., 2017). Such differences suggest that changes in fire frequency may not be spatially uniform across ecosystems (Baker & Williams, 2018). Potential geographic differences in fire frequency may arise if, for example, particular regions have been subject to lower rainfall or higher temperature deviation (relative to long-term mean values).

We sought to quantify patterns of geographic variation in fire frequency in the Australian State of Victoria, in the southeastern part of the continent. Victoria has been subject to several major wildfires over the past few decades (Collins et al., 2021; Enright et al., 2015; Lindenmayer & Taylor, 2020), and understanding patterns of fire frequency is important to assessing risks of ecosystem collapse (Bowman et al., 2014; Lindenmayer, Bowd, Taylor, et al., 2022) as well as determining resource availability for business sectors such as the forest industry (Cary et al., 2021). For our study in Victoria, we focused on the frequency of fires in each of two 20-year periods over the four decades between 1981 and 2020. In addition to our investigation being long term, it also was conducted over a large spatial scale—encompassing an area of 4.64 million ha, of which 4.16 million ha is primarily forest. The first question we posed was: Are there differences in the frequency of fires between the two 20-year time periods examined in this study? We postulated that the frequency of wildfires would be higher in the more recent 2001–2020 period relative to the preceding 20-year period. We made this prediction, in part, because of long-term warming and drying climatic conditions associated with climate change in southeastern Australia that can influence fire behavior (e.g., Canadell et al., 2021; Collins et al., 2022; van Oldenborgh et al., 2021). In addressing this question, we also sought to determine whether there was regional variation in fire frequency during our two time periods. To do this, we employed the Interim Biogeographic Regionalisation for Australia (IBRA), a regional classification developed for terrestrial and marine landscapes in Australia (and its external territories) (Environment Australia, 2000). There are 89 biogeographic regions and 419 subregions under the IBRA classification system. Each of these regions and subregions is distinguished by a unique set of key environmental attributes that, in turn, affect biodiversity and ecological processes (https://datasets.seed.nsw.gov.au/dataset/interim-biogeographic-regionalisation-for-australia-ibra-version-7-regions) (see Appendix S1). We predicted there would be substantial differences in fire frequency between IBRA subregions given that each ecosystem within these regions will likely be characterized by its own fire regime (Bradstock, 2010; Gill, 1975; Whelan, 1995) and this will be, in part, linked with differences in productivity and rainfall (Pausas & Ribeiro, 2013), and the composition of biotic communities (Rainsford et al., 2020).

The second phase of our work entailed addressing the question: What climatic and environmental factors are associated with fire frequency? To answer this question, we constructed statistical models to quantify the relationships between fire frequency and key climatic variables (temperature and rainfall), and potential environmental drivers such as slope, aspect, and elevation. We constructed a separate model for each of our six IBRA subregions and each of our two time periods. Potential geographic differences may arise if, for example, particular IBRA subregions have been subject to a greater rainfall deficit or higher temperature deviation (relative to long-term mean values), and therefore have experienced more fires than other regions where reductions in rainfall or increases in temperature have been less pronounced.

At the outset of our investigation, we made several a priori predictions in relation to the relationships between our potential explanatory variables and the number of fires recorded between 1981 and 2020 in each IBRA subregion.

Long-term climate effects: We predicted fire frequency would increase in those areas subject to large reductions in rainfall and large increases in long-term monthly average maximum temperature relative to long-term averages.

Elevation effects: We predicted that areas at higher elevation would experience more fires than locations at lower elevation, in part, because of the prevalence of lightning strikes (Kilinc & Beringer, 2007; Kotroni & Lagouvardos, 2008).

Slope effects: We predicted there would be more fires on steep slopes than on flat terrain as slope influences fire spread in dissected landscapes (Catchpole, 2002; Collins et al., 2019).

Aspect effects: We predicted that areas on drier, more northerly and westerly aspects will burn more frequently than locations on cooler and wetter southerly, easterly aspects (Catchpole, 2002).

Improved understanding of the frequency with which fires are occurring is essential for a range of key reasons. These include the impacts of fire frequency on: (1) fire-sensitive animal populations (Kelly et al., 2020; Ward et al., 2020; Whelan, 1995); (2) key parts of plant life cycles (Mahood & Balch, 2019) such as the ability to naturally regenerate (Day et al., 2020; Enright et al., 2015; Keeley & Pausas, 2019; Le Breton et al., 2022), and the ability to continue to successfully resprout following burning (Fairman et al., 2019); and (3) the potential to trigger regime shifts (Flores & Holmgren, 2021) such as through increasing flammability (Zylstra, 2018; Zylstra et al., 2022), impairing forest resilience (Hart et al., 2018), and in some cases resulting in ecosystem collapse (Bergstrom et al., 2021; Lindenmayer et al., 2011). Understanding the effects of fire frequency as part of fire regimes is also critical to determine the impacts of repeated fires on resource availability like supplies of timber (Cary et al., 2021; Cyr et al., 2009; Levine et al., 2022). In addition, understanding where fire is most likely to occur often (e.g., see Anderegg et al., 2022) is critical for guiding where to direct suppression efforts (Yebra et al., 2021). On this basis, new understanding from the empirical analyses presented in this study should be useful in helping determine where fire is occurring most frequently and how management actions might be tailored to meet the challenges that frequent, and often increasingly severe wildfires may create (see Collins et al., 2021).

METHODS Study area

Our study area consisted of forests within six IBRA subregions across eastern Victoria in southeastern Australia (Figure 1). We extracted data from these subregions from the IBRA7 dataset, which classifies Australia's landscapes into 89 large geographically distinct bioregions based on common climate, geology, landform, native vegetation cover, and other species information (Department of Climate Change, Energy, the Environment and Water, 2021). There are 419 IBRA subregions across Australia with more localized and homogenous geomorphological units in each bioregion (Department of Climate Change, Energy, the Environment and Water, 2021). We selected IBRA subregions that had >50% of their respective areas covered by a mapped Ecological Vegetation Class (EVC) Group in Victoria (DELWP, 2022b). The IBRA subregions we targeted were: (1) East Gippsland Lowlands, (2) Highlands-Northern Fall, (3) Highlands-Southern Fall, (4) Snowy Mountains, (5) South East Coastal Ranges, and (6) Victorian Alps. Collectively, these subregions cover an area of 4.64 million ha, of which 4.16 million ha is forest mapped under a given EVC Group (Appendix S2: Table S1). We further describe each of our six IBRA regions in Appendix S2: Table S1 and present summary data on them in Table 1. Around 3.8 million ha, equating to over 80% of the study area, is under public land tenure (Appendix S2: Table S2).

FIGURE 1. Map of the Interim Biogeographic Regionalisation for Australia (IBRA) regions in Victoria. We targeted six subregions for analysis: East Gippsland Lowlands, Snowy Mountains, South East Coast Ranges, Victorian Alps, Highlands-Southern Fall, and Highlands-Northern Fall. Further description of each subregion is provided in Table 1 and Appendix S1.

TABLE 1 Interim Biogeographic Regionalisation for Australia (IBRA) subregions, corresponding ecosystems listed in Bradstock (2010), Ecological Vegetation Class Groups and predominant ignition types.

IBRA subregion	Predominant ecosystems listed in Bradstock (2010) and likely matching Ecological Vegetation Class Groups used in this analysis	Fire season	Predominant ignition sources
East Gippsland Lowlands	WSF—Cool Temperate Wet Sclerophyll Forests (Wet or Damp Forests EVC Group; Lowland Forests EVC Group)	Late summer–autumn	Lightning
	DSF—Temperate Dry Sclerophyll Forests (Dry Forests EVC Group)	Spring–early summer	Human, variable lightning
Highlands-Northern Fall	WSF—Cool Temperate Wet Sclerophyll Forests (Wet or Damp Forest EVC Group)	Late summer–autumn	Lightning
	DSF—Temperate Dry Sclerophyll Forests (Dry Forest EVC Group)	Spring–early summer	Human, variable lightning
Highlands-Southern Fall	WSF—Cool Temperate Wet Sclerophyll Forest	Late summer–autumn	Lightning
	DSF—Temperate Dry Sclerophyll Forests (Dry Forests EVC Group)	Spring–early summer	Human, variable lightning
Victorian Alps	WSF—Cool Temperate Wet Sclerophyll Forest (Wet Forests EVC Group)	Late summer–autumn	Lightning
	DSF—Temperate Dry Sclerophyll Forests (Dry Forests EVC Group)	Spring–early summer	Human, variable lightning
	Treeless Subalpine Mosaic (Subalpine Grasslands, Shrublands or Woodlands EVC Group)	N/A	N/A
Snowy Mountains	WSF—Cool Temperate Wet Sclerophyll Forest (Wet Forests EVC Group)	Late summer–autumn	Lightning
	Treeless Alpine Mosaic (Subalpine Grasslands, Shrublands or Woodlands EVC Group)	N/A	N/A
South East Coastal Ranges	WSF—Cool Temperate Wet Sclerophyll Forest (Wet Forests EVC Group)	Late summer–autumn	Lightning
	DSF—Temperate Dry Sclerophyll Forests (Dry Forests EVC Group)	Spring–early summer	Human, variable lightning

Note: The typical “natural” inter-fire interval for all IBRA subregions is more than 20 years (see Bradstock, 2010). Further descriptions of IBRA subregions and their associated fire regimes are provided in Appendix S1.

Datasets used in analyses

We mapped areas subject to wildfires and planned burns that occurred between 1981 and 2020 using the Fire History Records of Fires across Victoria showing the fire scars dataset (DELWP, 2022a). This dataset represents the spatial extent of fires (inclusive of all fire severity classes) recorded primarily on public land since 1903. We extracted the spatial extent of areas burned by wildfires and planned burns separately for each year within a given EVC Group (DELWP, 2022b). We excluded areas burned by planned fires (e.g., prescribed [hazard reduction] burns) in each year. We computed the area burned by two or more fires by calculating the areas of overlapping fire polygons (Figure 2).

FIGURE 2. Extent and frequency of burned native vegetation (Ecological Vegetation Class Group) areas by wildfires between 1981 and 2000 (top) and 2001 and 2020 (bottom). IBRA, Interim Biogeographic Regionalisation for Australia.

To generate our topographic data, we used the Vicmap Elevation DEM 10 m Digital Elevation Model raster (DELWP, 2022c). It has a spatial resolution of 10 m with respective horizontal accuracy of 12.5 m and vertical accuracy of 5 m or better. From this, we generated slope and aspect raster layers in degrees (Appendix S3: Figure S2). We elected not to include other covariates for our statistical analyses (see below) such as a heat load index or transformed aspect. This was because such additional covariates would have been broadly correlated with other variables and created more complex models that would have been difficult to interpret.

We used historical climate data for our analysis. These data consisted of 30-year average baseline climate variables for the years 1951–1980. We followed the World Meteorological Organization's guidelines for climate normal data calculation for this period (World Meteorological Organization, 2017) as well as for the period of our study (1981–2020). We obtained gridded data for daily rainfall, and maximum temperatures from the Australian Gridded Climate Data (AGCD)/Australian Water Availability Project (AWAP) (Bureau of Meteorology, 2021). The spatial resolution of these gridded climate datasets was 0.05 × 0.05° (approximately 5 × 5 km) and the data accounted for spatial and temporal gaps in observations (Jones et al., 2009). We calculated the annual mean rainfall, and maximum and minimum temperatures for the baseline period of 1950–1979. We then calculated the total rainfall and the mean maximum and minimum temperature for each year between 1981 and 2020. We also calculated the annual difference for each year relative to the baseline mean for the years between 1981 and 2020. We grouped data from this study period into two subperiods for years between 1981–2000 and 2001–2020, respectively (Figure 3). We took this approach and used two distinct time periods because they encompassed times when robust climate and other data were available. In addition, the use of multiple decades and multiple years within decades meant that our analyses were not unduly influenced by exceptional fire behavior years.

FIGURE 3. Summary of climate data across the Interim Biogeographic Regionalisation for Australia (IBRA) subregions in our study detailing the mean annual baseline climate of 1951–1980 and the mean differences between the periods of 1981–2000 and 2001–2020. max temp, maximum temperature.

Data extraction

We sampled data using a square grid of points at 1000-m intervals across our six IBRA subregions. We extracted data on the number of wildfires at each point. We excluded points characterized by cleared land and which had not been assigned to an EVC Group. We also excluded points that had been subject to clear-cut logging or planned burns between 1981 and 2020 to focus our analysis on sites not directly impacted by intensive and/or recent anthropogenic disturbance, such as clear-cut logging and planned burns. Recent clear-cut logged sites have been observed to lead to increase fire severity, particularly under extreme weather conditions (Lindenmayer et al., 2021; Taylor et al., 2014). Previous planned burns may increase fire severity depending on the time since the planned burn (Zylstra et al., 2023). The exclusion of disturbed sites by clear-cut logging and/or planned burns resulted in 19,563 points for analysis across the study area (Table 2).

TABLE 2 Analysis points across the Interim Biogeographic Regionalisation for Australia (IBRA) subregions across the study region.

IBRA subregion	No. sites
East Gippsland Lowlands	1569
Highlands-Northern Fall	4713
Highlands-Southern Fall	5469
Snowy Mountains	1077
South East Coastal Ranges	2835
Victorian Alps	3900
Total	19,563

STATISTICAL METHODS

We used Poisson generalized additive models (GAMs) (Wood, 2017) to address our two key research questions on associations between fire frequency and IBRA subregion, time period, topographic features, climate deviations, and smooth functions of longitude and latitude. We describe our models more fully below. Let $$$ {y}_i\left({\boldsymbol{s}}_{\boldsymbol{i}}\right) $$$ represent the number of fires experienced at the geo-referenced point $$$ {\boldsymbol{s}}_{\boldsymbol{i}} $$$, where $$$ {\boldsymbol{s}}_{\boldsymbol{i}} $$$ is a two-dimensional vector of spatial coordinates: longitude and latitude. Let $$$ {\boldsymbol{x}}_{\boldsymbol{i}}\left({\boldsymbol{s}}_{\boldsymbol{i}}\right) $$$ represent the vector of covariates available at the geo-referenced point $$$ {\boldsymbol{s}}_{\boldsymbol{i}} $$$. Our Poisson GAM model is given by the following:[Image Omitted. See PDF]where $$$ \boldsymbol{\upbeta} $$$ is the vector of unknown regression coefficients, $$$ \boldsymbol{s}\left({\mathrm{long}}_i,{\mathrm{lat}}_i\right) $$$ is the generalized additive component or smooth term of the model (Gaussian process), which we allowed a maximum degrees of freedom of 200, and T indicates matrix/vector transpose. We fit the models using the mgcv package (Wood, 2011) in R version 4.1.2 (R Core Team, 2021) using the maximum likelihood. We describe the form of the covariate vector $$$ {\boldsymbol{x}}_{\boldsymbol{i}}\left({\boldsymbol{s}}_{\boldsymbol{i}}\right) $$$ for our two questions in more detail below.

Spatial confounding can have drastic effects on the interpretation of the covariates in a spatial regression model (Hodges & Reich, 2010). Hodges and Reich (2010) recommend the use of restricted spatial regression (RSR), which constructs the spatial random effects (smooth terms) orthogonal to the fixed effects. In the Gaussian (identity link) case, the RSR regression coefficients are identical to the nonspatial regression coefficients with inflated standard errors to account for the lack of independence in spatial data. In the Poisson regression case, the RSR regression coefficients are consistent with the nonspatial model but are not identical. We report RSR regression coefficients.

To avoid spatiotemporal complications, we addressed our first question by fitting a Poisson GAM to each of our two time periods (1981–2000 and 2001–2020) with the covariate vector, $$$ {\boldsymbol{x}}_{\boldsymbol{i}}\left({\boldsymbol{s}}_{\boldsymbol{i}}\right) $$$, consisting of indicator (dummy) variables for each of the six IBRA subregions. The covariate vector for our second question on the effects of climatic and environmental factors on fire frequency consisted of the aspect (easting and northing), elevation (DEM, in meters above sea level), slope (in degrees), the relative change in average maximum temperature in each period compared with the historical period (1950–1979), and the relative change in average rainfall. We fit these models for each IBRA subregion and for each time period.

We employed a permutation test to determine whether there were differences in the frequency of fires between the first period (1981–2000) and the second period (2001–2020). For each point, we randomly assigned the observed total number of fires to each time period. We then repeated this for all points, then computed the difference between the two time periods. We did this 10,000 times, and this formed the reference distribution to test the null hypothesis that there was no difference in fire frequency between our two 20-year time periods. We then compared the observed values to the reference distribution. We completed residual diagnostics for each of the fitted models to assess the assumption and computed and plotted empirical variograms for each time period and IBRA region combination.

RESULTS Q1: Are there differences in fire frequency between the two 20-year time periods? Do temporal differences in fire frequency vary between IBRA subregions?

Between 1981 and 2000, 666,704 ha of forest were burned, with 36,050 ha being burned more than once during that time. Between 2001 and 2020, 3.1 million ha burned, with 1.03 million ha burned more than once in that time (Appendix S3: Figure S1). The total number of fires that burned between 1981 and 2020 at a given point within the six IBRA subregions varied from zero to five (Figure 4). Between 1981 and 2000, the dominant fire frequency category for all IBRA subregions, except the East Gippsland Lowlands was zero fires (Figure 4).

FIGURE 4. Frequency of fires for two 20-year periods between 1981 and 2020. Magenta bars show values for the frequency of fire at a given point between 1981 and 2000. Yellow bars correspond to the frequency of fires at a given point between 2001 and 2020.

Our data indicated that in all six IBRA subregions, there were a greater number of fires between 2001 and 2020 than in the preceding 20-year period (1981–2000) (Figure 5). A permutation test revealed that the null hypothesis predicting no difference in fire frequency between time periods should be rejected (Appendix S3: Figure S3). The change in the number of fires between time periods was most pronounced in three IBRA subregions in particular: Snowy Mountains, East Victorian Alps, and South East Coast Ranges (Figure 5). The least amount of change was in the East Gippsland Lowlands, which had the greatest number of fires in 1981–2000 relative to all other IBRA subregions, but an intermediate number of fires in the second time period (Figures 4 and 5). An analysis of 15 cross-IBRA comparisons showed there were differences in fire frequency between all IBRA subregions, both in period 1 and period 2 (Appendix S3: Figure S4).

FIGURE 5. Estimated number of fires in different Interim Biogeographic Regionalisation for Australia (IBRA) subregions in two 20-year time periods (1981–2000 shown in magenta, 2001–2020 shown in yellow). Differences between the two time periods are significant in all IBRA subregions (see Appendix S3: Figure S3).

Q2: What factors are associated with fire frequency within different IBRA subregions?

We quantified the factors associated with the frequency of fires in each IBRA subregion and in each time period (Figure 6; Appendix S3: Figure S3). The low number of fires in the first time period (1981–2000) meant there was limited variation in our data for the response variable, with only the East Gippsland Lowlands experiencing many fires in that time (see Appendix S3: Figure S5). On this basis, we have not presented the results of analyses of covariate effects for our first time period.

FIGURE 6. Covariate associations with fire frequency in different Interim Biogeographic Regionalisation for Australia subregions for the period from 2000 to 2020. Note that the covariates have been standardized to have a zero mean and standard deviation of one prior to analyses.

In the second time period (2001–2020), fire frequency increased with increasing rainfall deficit deviation in four IBRA subregions (South East Coastal Ranges and East Gippsland Lowlands), but the opposite trend characterized the Victorian Alps and Highlands—Southern Fall (Figure 6). We found that associations between fire frequency and increasing temperature deviation varied from negative to positive depending on IBRA subregion (Figure 6). There was a negative relationship in the Highlands-Southern Fall and Highlands-Northern Fall, but the opposite pattern (more fires with increasing temperature deviation) in the East Gippsland Lowlands and the Snowy Mountains (Figure 6). There was further complexity in the contrasting rainfall deficit and temperature deviation effects. In some IBRA regions, the relationship with fire frequency was positive for both rainfall and temperature measures (e.g., Snowy Mountains, East Gippsland Lowlands), in others it was negative for both measures (Highlands-Southern Fall), and yet in others there was a positive or negative response to rainfall deficit (Victorian Alps, South East Coastal Ranges) but no response to temperature deviation (Figure 6).

Our analyses revealed complex associations between fire frequency and aspect, which varied from positive for northing to negative for easting (Victorian Alps), positive to northing only (Highlands-Northern Fall), and positive for both northing and easting (Highlands-Southern Fall) (Figure 6). We found that the number of fires was more frequent at higher elevations in five of the six IBRA subregions, with the most marked effects in the Victorian Alps IBRA subregion (Figure 6). In the case of slope, relationships with fire frequency ranged from marginally negative (South East Coastal Ranges), to positive in all other IBRA subregions except the East Gippsland Lowlands (where there was no slope effect) (Figure 6).

We present the diagnostics for our RSR models in Appendix S3: Figures S6–S17 and they show that the spatial model fitted the data better than the nonspatial model. This indicated the importance of accounting for spatial dependence in our analyses of fire frequency.

DISCUSSION

Altered fire regimes pose a major risk to ecosystems (Flores & Holmgren, 2021; Syphard et al., 2022), biodiversity (Keeley & Pausas, 2019; Kelly et al., 2020), and human life and property (Bushfire Royal Commission, 2020; Gibbons et al., 2012; Moritz et al., 2014). Understanding how they might be changing is critical to making informed decisions about both fire management and ecosystem management (Levine et al., 2022; Santos et al., 2022). Our empirical analyses revealed that fire frequency in Victoria was heterogeneous both in time and in space as indicated by evidence of: (1) a change in the frequency of wildfires over the past 40 years (between 1981 and 2020), with a marked increase in the past 20 years (2001–2020) relative to the 20 years prior (1981–2000); (2) changes in fire frequency not being spatially uniform, with increases more pronounced in some IBRA subregions than others; and (3) several climate and topographic factors influencing the frequency of wildfires, but their effects manifest varyingly in different IBRA subregions. We discuss these key findings in more detail below including their implications for fire and ecosystem management.

Q1: Is fire frequency increasing?

We found compelling evidence of recent increases in fire frequency in Victoria (Figures 4 and 5). There was a significant increase in the period 2001–2020 relative to the period 1981–2000. Our results are broadly consistent with those predicted to occur over large spatial scales as a result of climate change (Canadell et al., 2021; Jones et al., 2020; Westerling et al., 2006), although as outlined below, we found highly variable responses between IBRA subregions to rainfall deficit and temperature deviation (Figure 6).

We also found that changes in the frequency of fires were more pronounced in some IBRA subregions than others, highlighting patterns of spatial variability that have not previously been documented. The reasons for such between-IBRA subregional differences in fire frequency remain unclear. However, different IBRA subregions are, by definition, characterized by different bioclimatic, geomorphological, and other attributes (Environment Australia, 2000), and these would be expected to influence key aspects of fire regimes (Bradstock, 2010) including the frequency of fire. For example, increasing temperature may have increased the number of days in which fuels were dry enough for fire to propagate in otherwise cool or wet regions, or driven a reduction in ephemeral grass cover in drier, hotter regions. In particular, we note that changes in fire frequency between periods were most marked in those IBRA subregions at generally higher elevations such as the Snowy Mountains, Victorian Alps, and South East Coastal Ranges (Figure 4; Appendix S3: Figure S3). The IBRA subregion at the lowest elevation—the East Gippsland Lowlands—exhibited the least amount of change between time periods, possibly because fire frequency was already relatively high there in our first time period 1981–2000 (Figures 4 and 5). Elevation was a key explanatory variable in models for each of the IBRA subregions (see Figure 6) except for the Snowy Mountains where fire frequency was greater in the second period relative to all other subregions. This suggests a complex array of factors (including some not measured in this study) may have an important influence on fire frequency in our study area (see below). We also note that data availability meant that over 80% of our study area falls within public land tenure and this may have biased our results in terms of analyzing different land tenures, although in presently unknown ways.

Q2: What factors are associated with fire frequency within different IBRA subregions?

We limited our analyses of covariate effects to our second 20-year period (2001–2020). This was because fires were relatively infrequent in the initial 20-year period resulting in limited variation in our data. Our analysis indicated that a complex interplay of climatic and environmental factors influenced fire frequency, with the effects of these factors varying between IBRA subregions. For example, fire frequency increased with increasing rainfall deviation in four of the six IBRA subregions in the second time period, exceptions were the Highlands-Southern Fall and the Victorian Alps where it declined with increasing rainfall deficit (Figure 6). There also were IBRA subregion differences in fire frequency that ranged from negative to positive for temperature deviation relative to the long-term average (Figure 6). These rainfall and temperature effects were more complex and spatially nuanced than we anticipated at the outset of this study (see a priori predictions in the Introduction). Indeed, the relationships we quantified conformed our predicted responses for only some of the IBRA subregions we analyzed (see Figure 6). Thus, while climate change may have an important influence on fire frequency overall, more subtle and variable effects of factors associated with climate change like rainfall deficit and temperature deviation are also important at the IBRA subregion level. For example, increased mean rainfall may in some cases arrive in storm events with associated lightning, and reduced rainfall retention due to faster runoff.

We found evidence of elevation effects, with more frequent fires at higher elevation within most IBRA subregions (Figure 6). This is consistent with global analyses showing that high mountain areas in particular are subject to an increasing amount of wildfire (Xu & You, 2022). Several factors may explain this result. First, locations at higher elevations are a greater risk of lightning strikes (leading to more fires), than places positioned lower in the landscape (Cattau et al., 2020; Kotroni & Lagouvardos, 2008). Indeed, elevation effects were more pronounced in the Victorian Alps (where much of the alpine and subalpine environment occurs in Victoria) than in other IBRA subregions (Figure 6). Second, places at high elevation may generally also be steeper, with fires sometimes more likely to burn on steep slopes such as through spotting (see Storey et al., 2020), as indicated by the slope effects we identified for most IBRA subregions in the second time period (Figure 6). Third, warming and drying at high elevations also may be a factor contributing to our results, although data are currently lacking on the sources of ignitions for fires over the past 40 years in our study area.

IMPLICATIONS FOR MANAGEMENT

Our investigation revealed that fire frequency was much higher in the second period (2001–2020) than the first time period and for all IBRA subregions in Victoria. This outcome was expected given the changes in climate that have been documented in the State over the past decades (Cai & Cowan, 2008; Canadell et al., 2021). The particularly pronounced change in fire frequency in the past two decades is a cause for major concern, especially in those IBRA subregions where wildfires have become particularly prevalent. For example, parts of those IBRA subregions, which support wetter forest types like those dominated by ash-type eucalypts, have been subject to many wildfires (Enright et al., 2015; Lindenmayer & Taylor, 2020). These forests are also where the majority of Victoria's native forest logging operations take place (Lindenmayer & Taylor, 2022). However, the short fire return interval in these ecosystems, coupled with the prolonged fire-free period needed for trees to grow to age where they can be cut for viable sawlogs, means that forestry industries based on this form of wood production are in a precarious state (Cary et al., 2021). An increase in fire frequency (and a corresponding contraction in the interval between fires) will mean considerable uncertainty in sawlog supplies from these kinds of native forests (Cary et al., 2021). This issue is magnified by the fact that logged and then regenerated forests are also more flammable for a period of 40–70 years after they were last harvested (Taylor et al., 2014; Wilson et al., 2022).

Some of the ecosystems within the IBRA subregions that we analyzed have experienced greater frequencies of fires than would be “natural” under “past fire regimes” (sensu Bradstock, 2010). This is, for example, the case of alpine woodlands and montane forests (e.g., see Banks, 1982; McCarthy et al., 1999; Zylstra, 2018). Such changes in fire regimes can result in forests becoming degraded (Fairman et al., 2019) or even being at risk of ecosystem collapse (Bergstrom et al., 2021; Enright et al., 2015; Lindenmayer, Bowd, Taylor, et al., 2022). A key management issue is how to limit the number of wildfires in these areas. Given that climate and fire weather are two of the key drivers underpinning wildfire, an obvious response is to embrace more strident steps to tackle climate change. However, this is a huge global challenge and there will be a prolonged lag time between actions taken now and demonstrable changes in climatic conditions (IPCC, 2023). More immediate options include vegetation management actions that reduce flammability such as limiting logging in wood production forests and precluding hazard reduction burning in those vegetation types where such activities can sometimes have perverse effects and make them more (not less) flammable (Campbell et al., 2022; Zylstra, 2018; Zylstra et al., 2022).

We argue that new technologies and thinking are needed to support better fire management (Roldán-Gómez et al., 2021; Yebra et al., 2021). Early detection of wildfires is critical, particularly on extreme fire risk days when fires can have catastrophic impacts (Lindenmayer, Zylstra, et al., 2022). The earlier that a fire is detected, the smaller the fire will be when fire crews first arrive, and the greater the likelihood that the fire can be successfully contained and suppressed at this early stage of fire development (Plucinski, 2012). These are consistent with “ecologically cooperative” approaches to fire that reinforce natural controls on flammability to create less flammable and more fire-resilient forest landscapes (Zylstra et al., 2022, 2023).

CONCLUSIONS

Fire frequency is a key component of fire regimes globally. Changes in fire regimes can have profound impacts on biodiversity (Kelly et al., 2020) and on the condition of ecosystems, including elevating the risks of ecosystem collapse (Lindenmayer, Bowd, Taylor, et al., 2022). Australia has been widely regarded as one of the most fire-prone jurisdictions on Earth, with recent conflagrations in the eastern part of the continent in 2019–2020 found to be unparalleled in terms of total area burned, the extent of high-severity wildfire (Collins et al., 2021), the uniformity of burned areas (lacking patchy unburned refugia) (Mackey et al., 2021), and the amount of biodiversity lost (Ward et al., 2020). A deeper understanding of fire impacts demands the quantification of key aspects of multiple fires over time in a given area or ecosystem. The analyses presented here show that for the Australian State of Victoria, there has been a marked increase in the frequency of wildfires in the past 20 years relative to the 20 years preceding that. These increases have been spatially variable, and more pronounced in some IBRA subregions than others. The frequencies of fires were strongly associated with some key covariates including rainfall and temperature deviance from the long-term mean, elevation, and slope. However, how these effects were manifested in varied ways between IBRA subregions.

ACKNOWLEDGMENTS

We acknowledge the many First Nations in Victoria and their Elders past and present on whose respective Countries our study was conducted. We also acknowledge their Custodianship is essential to conserving the forests and ecosystems analyzed in our study. We thank the Victorian Government, and the Bureau of Meteorology for access to key datasets that enabled the analyses reported here to be completed. Tabitha Boyer and Luke Gordon assisted with many key aspects of manuscript preparation. Comments from two anonymous referees greatly improved an earlier version of the manuscript.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

DATA AVAILABILITY STATEMENT

Data are available from the Bureau of Meteorology (2021), DELWP (2022a, 2022b, 2022c, 2022d), and the Department of Climate Change, Energy, the Environment and Water (2021).