Pyrogeography in flux: Reorganization of Australian fire regimes in a hotter world

Changes to the spatiotemporal patterns of wildfire are having profound implications for ecosystems and society globally, but we have limited understanding of the extent to which fire regimes will reorganize in a warming world. While predicting regime shifts remains challenging because of complex climate–vegetation–fire feedbacks, understanding the climate niches of fire regimes provides a simple way to identify locations most at risk of regime change. Using globally available satellite datasets, we constructed 14 metrics describing the spatiotemporal dimensions of fire and then delineated Australia's pyroregions—the geographic area encapsulating a broad fire regime. Cluster analysis revealed 18 pyroregions, notably including the (1) high‐intensity, infrequent fires of the temperate forests, (2) high‐frequency, smaller fires of the tropical savanna, and (3) low‐intensity, diurnal, human‐engineered fires of the agricultural zones. To inform the risk of regime shifts, we identified locations where the climate under three CMIP6 scenarios is projected to shift (i) beyond each pyroregion's historical climate niche, and (ii) into climate space that is novel to the Australian continent. Under middle‐of‐the‐road climate projections (SSP2‐4.5), an average of 65% of the extent of the pyroregions occurred beyond their historical climate niches by 2081–2100. Further, 52% of pyroregion extents, on average, were projected to occur in climate space without present‐day analogues on the Australian continent, implying high risk of shifting to states that also lack present‐day counterparts. Pyroregions in tropical and hot‐arid climates were most at risk of shifting into both locally and continentally novel climate space because (i) their niches are narrower than southern temperate pyroregions, and (ii) their already‐hot climates lead to earlier departure from present‐day climate space. Such a shift implies widespread risk of regime shifts and the emergence of no‐analogue fire regimes. Our approach can be applied to other regions to assess vulnerability to rapid fire regime change.


| INTRODUC TI ON
Fire has played a crucial role in the Earth system for millions of years resulting in metastable climate-vegetation coupling known as fire regimes (Bowman et al., 2009).Changes to fire regimesthe prevailing spatiotemporal characteristics of fire-are imposing heavy costs on ecosystems (Kelly et al., 2020) and societies (Filkov et al., 2020).These changes include global increases in extreme fire activity (Collins et al., 2021;Turco et al., 2023;Whitman et al., 2022), driven by a drying trend of the atmosphere and fuel (Ellis et al., 2022;Jain et al., 2022) and intensified fire weather (Canadell et al., 2021;Jolly et al., 2015;Jones et al., 2022).Australia's Black Summer bushfires of 2019/2020, for example, occurred during the nation's hottest and driest year on record, and were consequently unprecedented in their scale, severity, and impacts (Abram et al., 2021;Boer et al., 2020;Collins et al., 2021;Nolan et al., 2020).Similarly intense fire seasons have occurred in recent years in Brazil, Canada, western United States, Europe, Indonesia, Chile, and Siberia (Bowman et al., 2020;Linley et al., 2022), signaling the possible emergence of novel fire regimes across Earth (Bowman et al., 2021).
Climate change will influence fire regimes through its effects on each of four limiting factors on landscape fire-biomass, dryness, ignitions and weather-referred to as 'switches' by Bradstock (2010).
For instance, grassy savanna promotes fire, which can prevent tree incursion, while canopy closure suppresses fire, in turn maintaining closure (Murphy & Bowman, 2012;Staver et al., 2011b).Grazing can switch grassy ecosystems to woody ones with marked changes in fire regimes (Williamson et al., 2016).State change can be incremental or abrupt (Bergstrom et al., 2021;Harris et al., 2018) and may involve shifts in the key limitations on fire, such as shifting from being limited by fuel dryness to fuel productivity (Boer et al., 2016).Feedbacks and the sometimes-abrupt nature of regime shifts are further complicated by anthropogenic influences that affect ignitions, fuel types and fire spread (Balch et al., 2017), including loss of ancient traditions of Indigenous fire management (Bowman et al., 2022).Despite these challenges in predicting the exact nature of regime shifts, it is critical to anticipate locations most at risk of regime shifts.
Identifying the climate niches of fire regimes provides a means of evaluating the risk of regime shifts.Just as species occur within distinct multidimensional environmental niches (Holt, 2009), so too do fire regimes (Krawchuk et al., 2009;Moritz et al., 2010Moritz et al., , 2012)).Thus, evaluating whether projected climates occur within a regime's climate niche should provide an indication of the risk of a regime shift.A future climate without modern counterparts, for example, might be expected to yield a fire regime that also lacks modern counterparts, variably referred to as novel or no-analogue ecosystems (Williams & Jackson, 2007).Identifying these climate niches first requires mapping the geographical unit of a fire regime, for which there have been numerous approaches ranging from global pyromes (analogous to biomes; Archibald et al., 2013) to finer-scale pyroregions (analogous to ecoregions; Elia et al., 2022).Prior global analyses have identified 5-15 coarse global pyromes based on satellite data (Archibald et al., 2013;Chuvieco et al., 2008;Pais et al., 2023), while Murphy et al.'s (2013) finer-scale analysis identified 20 Australian fire regimes using expert knowledge and vegetation maps.Despite these studies, there remains significant scope to improve our empirical understanding of Australia's fire regimes (and indeed the world's), their climate niches, and the extent to which they might reorganize under climate change.
Australia's broad climate gradients, diversity of fire-prone ecosystems, and the ancient cultural use of fire, make it an invaluable pyrogeographic laboratory (Bowman & Murphy, 2011).The most fire-prone regions of Australia, the northern monsoonal savannas and the southern temperate forests, have discrete intra-annual periods of wetness and dryness (Williamson et al., 2016), yielding abundant fuel that is periodically available to burn.Natural inter-annual climate variability can further promote fire (Abram et al., 2021;Williamson et al., 2016).Australia's Aboriginal peoples have for millennia used low-intensity patchy fire to promote biodiversity, reduce fuel loads, and maintain healthy country (Adeleye et al., 2021;Steffensen, 2020).This use of fire involves reading landscapes to determine when human-ignited fire would generate desirable outcomes in different vegetation types (Steffensen, 2020).Multiple strong climate gradients, frequent fire, the legacy of ancient cultural use of fire, and the importance of fire in the nation's greenhouse gas accounts (Bowman et al., 2023;Haverd et al., 2013), not only make Australia an ideal pyrogeographic laboratory but also an immensely societally and ecologically consequential region in which to anticipate the effects of climate change on fire regimes.
In this study, we aim to delineate Australia's pyroregions-the geographic area surrounding internally consistent fire regimes-and then identify locations most susceptible to regime shifts.To do this, we used four decades of satellite observations to define a series of characteristics of fire activity (e.g., extent, shape, seasonality, intensity and interannual variability; Gill, 1975).Next, we used a clustering algorithm to delineate Australia's pyroregions.Finally, we analyzed the climate niches of these fire regimes and predicted locations most at risk of fire regime shifts.We infer the risk of a fire regime shift by identifying locations where the projected climate envelope is (i) novel to a given pyroregion, indicating a likely regime shift, and (ii) novel to the entire continent, indicating the likely emergence of fire regimes for which there are no present-day analogues (Williams & Jackson, 2007).
Our analysis involved five main steps summarized in Figure 1.All analyses were conducted in R version 4.3.0(R Core Team, 2023).

| Metrics of fire
We developed metrics that describe the spatiotemporal dimensions of fire regimes using satellite datasets described in Table 1.From the satellite products, we constructed 14 metrics of fire activity and fuel, with all analyses conducted at a spatial resolution of 0.25° (~25 km).This resolution was selected as a match for the continental scale of the analyses.The relatively coarse resolution also ensured that fire had been observed in most cells, helping to mitigate the issue of the relatively short satellite record (21-37 years) compared to the return intervals of fire in some systems (Murphy et al., 2013).
The 14 fire metrics included (1) daytime hotspot density, (2) the relative difference between day and night hotspots, (3) fire radiative power at 0.95th quantile, (4) week of peak fire activity, (5) fire season length, (6) median daily patch size of fires, (7) maximum daily patch size of fires, (8) pyrodiversity (richness of burn ages within a grid cell), ( 9) time since fire contagion (homogeneity of burn ages), (10) mean monthly greenhouse gas emissions, (11) median annual proportion burned, (12) interannual coefficient of variation of proportion burned, (13) mean net primary productivity (NPP) as an estimate of biomass accumulation, and an indirect estimate of fuel load, and (14) normalized difference vegetation index (NDVI) difference between summer and winter as a measure of the relative timing of plant growth and senescence (also known as curing ;Chaivaranont et al., 2018).The pyrogeographic significance and construction of these variables are described in Table 2.
Once constructed, all fire variables were normalized using the bestNormalize package in R, which selects the approach that best normalizes the data from a series of transformations (Peterson, 2021).In our case, all fire variables were transformed using an ordered quantile normalization (Peterson & Cavanaugh, 2020), and then rescaled from 0 to 1 to place them on a common scale.

| Regionalization of fire regimes
To classify Australia into clusters with distinct fire regimes (hereafter "pyroregions"), we fitted Gaussian mixture models using the mclust package in R (Scrucca et al., 2016).We fitted each of the 14 types of multivariate models offered by mclust, representing different parametrizations of the covariance matrix that constrains the shape, volume, and orientation of the clusters (Scrucca et al., 2016).We fitted these models with 1-30 clusters, and then used Bayesian information criterion (BIC; Fraley & Raftery, 1998) to identify the best-performing model type and the optimal number of clusters.The variable shape, variable volume, variable orientation (Scrucca et al., 2016) model clearly provided the best fit to the data.However, as in Archibald et al.'s (2013) global pyrome analysis, the addition of extra clusters led to diminishing and eventually negligible improvement in BIC.Thus, in pursuit of the simplest adequate model, we selected a model with 17 clusters, as models with 18 or 19 clusters did not improve model fit significantly, after which point BIC began to worsen (Figure S1).
Using the categorization from mclust, we produced a map of Australia's pyroregions.Because some areas in central Australia have not had fire recorded during the MODIS record (e.g., Figure 2e), some fire metrics could not be calculated, and these areas were therefore not included in the model.However, these areas in fact represent low-frequency regimes; for example, chenopod and mulga shrublands with sparse grass cover have fire return intervals of 20-100 years (Murphy et al., 2013) that exceed the length of the current satellite record.We therefore assigned these locations without MODIS observed fire to an additional post-hoc category representing a low-frequency regime (as opposed to retaining these locations as missing values).Finally, we used a 5 × 5 majority smoothing window to produce a smoothed map of Australia's pyroregions.As distinct from other clustering approaches, model-based clustering provides 'soft-assignment' by estimating the probability and uncertainty of assigning observations to clusters (Scrucca et al., 2016).We therefore interpreted our results in the context of this uncertainty (Figure 3a; Figure S2), which included producing the final map with regimes ordered from most to least certain.
To describe the key fire features and their strength of association with each pyroregion, we fitted a random forest model (Breiman, 2001) that predicted the pyroregion class in response to the 14 fire and fuel variables.The random forest was fitted with 1000 trees and validated with 10-fold cross-validation summarized with the Kappa statistic using the 'caret' package in R (Kuhn, 2008).
F I G U R E 1 Workflow summary of the key stages of the analysis.Briefly, satellite datasets were used to create metrics of fire regimes.Using those metrics, Australia was divided into pyroregions with internally consistent fire characteristics.For each pyroregion, we characterized the climate niches and the extent to which alternative climate projections occur within those niches.
Model accuracy was evaluated on 30% of data withheld from model fitting.We then summarized the class-specific variable importance (scaled to sum to 1 for each regime) based on mean decrease in the prediction error on the out-of-bag data while permuting predictor variables ('importance' function of the 'randomForest' package ;Liaw & Wiener, 2002).In addition to the strength of association, we also produced violin plots that demonstrate the direction and distribution of the fire metrics for each pyroregion (Figure S3).
To describe the land cover attributes of the pyroregions, we summarized proportional land cover of each pyroregion using the Australian Dynamic Land Cover Dataset version 2.1 (250 m resolution; McIntyre et al., 2015).This dataset consists of 22 Australian land cover classes that reflect the structural character of vegetation, ranging from crops and pastures to natural land covers like forests and grasslands.For simplicity, we aggregated irrigated and non-irrigated cropping, pasture, and sugar production into a single cropping/pasture category.For interpretability, we named the pyroregions based on a combination of climate zone, dominant land cover, and dominant vegetation types.

| Associations between pyroregions and climate
We characterized the historical climate (mean from 1970 to 2000) niche of Australia's pyroregions, and then evaluated the likely extent of change by 2081-2100 under three Shared Socioeconomic Pathways (SSP) of the Coupled Model Intercomparison Project 6 (CMIP6).SSP1-2.6 is the most optimistic scenario that we examined, representing a world that makes gradual but pervasive advances towards decarbonization (Riahi et al., 2017).SSP2-4.5 represents a middle-of-the-road scenario that does not deviate markedly from historical patterns, with some countries making good advances and others not (Riahi et al., 2017).SSP3-7.0 is a mid-range scenario for a world that fails to enact any climate policies (Riahi et al., 2017).SSP2-4.5 and SSP3-7.0 are expected to be among the most realistic scenarios (Hausfather & Peters, 2020).While frequently analyzed, we chose not to include SSP5-8.5 (formerly RCP8.5) because it was intended as an unlikely worst-case scenario and its frequent TA B L E 1 Data sources used to construct metrics that describe Australia's fire regimes.

Data source Data type and temporal scale Description
MODIS active fire records ("hotspots") 4 × daily point locations from 2001 to 2021, totalling 4.39-million points Point locations are provided at 1-km spatial resolution based on detection of thermal anomalies (MCD14ML; Giglio et al., 2018).Fire radiative power (FRP; megawatts) provides a means of distinguishing low-intensity (e.g., agricultural burn-offs) and highintensity fires (e.g., crown fires; Giglio et al., 2006).
Four satellite passes per day facilitate comparison of day and night fire activity (Giglio et al., 2006;Williamson et al., 2016) MODIS burned area polygons Daily polygons from 2000 to 2021, totalling 5.86-million polygons Polygons are derived at 500-m spatial resolution from spectral analysis and temporal change techniques (MCD64A1; Giglio et al., 2015).These polygons allow us to examine patterns in the shape and arrangement of the area burned by individual fires at daily increments, helping to characterize both the size of fires and their rate of progression The week of the year with the most hotspots.Because year is a circular variable, which clustering algorithms cannot accommodate, peak fire week was calculated as the absolute number of weeks from week 36 (early September), when the disparity between the northern and southern fire seasons was at its greatest

5.
Fire season length Temporal period over which fire occurs The minimum number of consecutive days required for a cell to record 90% of its total hotspot activity.This was calculated by iterating through each day of the year as the start of the fire season and calculating the number of days for a cell to go from 5% to 95% of its cumulative hotspot count.We calculated this metric at a resolution of 0.5° to reduce stochasticity in areas with infrequent fire, after which cell size was disaggregated back to 0.25°( ii) MODIS burned area 6.

Median daily patch area Spatial pattern of the daily area burned by individual fires
The area of each daily burned area polygon was rasterized at 0.01° resolution, such that all cells within a fire polygon were given the value corresponding to the total area burned that day.Next, the median daily patch area (km 2 ) was calculated for each cell, characterising the daily spread of a fire  through the interactive effects of temperature and precipitation on plant growth (Buermann et al., 2018), vapour pressure deficit (Ficklin & Novick, 2017;Jain et al., 2022), and its drying effect on soil and fuel (Nolan et al., 2016).
To synthesize the predictions from different Global Climate Models (GCMs) and reduce the effect of any single model's idiosyncrasies, we created an ensemble average of WorldClim's downscaled projections of the abovementioned five variables (https:// www. world clim.org/ data/ cmip6/ cmip6 clima te.html# ).We used the following five GCMs: ACCESS-CM2 (Bi et al., 2020) Prior to further analyses, we jointly scaled the historical and future climate variables between 0 and 1.
To characterize the strength of associations between pyroregions and the current climate, we fitted a random forest with pyroregion class modelled in response to the ensemble-averaged climate variables.Class-specific variable importance was used to provide a regime-specific measure of the relative strength of climate associations.Model fitting and variable importance followed the same approach as described in the previous section.
After demonstrating that the pyroregions could be predicted with high accuracy by these climate variables, we next characterized the extent to which future climates are projected to occur within each pyroregion's historical envelope.Because of the often abruptnature of regime shifts (Bergstrom et al., 2021;Harris et al., 2018) and the challenges imposed by climate-vegetation-fire feedbacks (Bowman et al., 2014), we chose to investigate the spatial extent of climate novelty rather than predict what a regime will shift to (which could either be an extant or entirely novel state).To do this, we characterized each pyroregion's climate niche using an n-dimensional hypervolume (Blonder et al., 2014(Blonder et al., , 2018)).The hypervolume R package (Blonder et al., 2018) provides two methods for estimating hypervolumes: (i) a Gaussian kernel density estimate provides a 'loose wrap' of the hypervolume, and (ii) a one-class support vector machine provides a 'tight wrap' of the hypervolume (Blonder et al., 2018).Based on both approaches, we first characterized the n-dimensional hypervolume for each regime under the historical climate.Next, we tested whether the projected climates of each cell were inside or outside of (i) the pyroregion's historical hypervolume and (ii) the hypervolume of any extant pyroregion ('hypervolume_in-clusion_test' function; Blonder et al., 2014).If the projected climate did not occur within the hypervolume of any other pyroregion, we considered the projected climate to be novel to the Australian continent.Finally, we quantified and visualized the agreement between the loose and tight wrap methods.

| Australia's pyroregions
There were clear regional differences in the patterns of fire across Australia (Figure 2).Hotspot density was highest in the monsoon tropics (Figure 2a), where relatively small fires (Figure 2c) burnt Figure 2b).As expected, intense fires were observed in the temperate forests of the south, but interestingly, very intense fires were also observed over large areas of the low-productivity central west (Figure 2e), where spinifex grasslands have low productivity but can accumulate large fuel loads over time.There was a clear north-south division in the timing of peak fire season, with hotspots peaking near the end of the dry season (spring) in the monsoon tropics, while they peaked late in the Austral summer/ autumn in southern Australia (Figure 3k).
The regional patterns of fire gave rise to 18 pyroregions (Figure 3), comprised of 17 pyroregions identified by the Gaussian mixture model (Figure S1) and one post-hoc pyroregion in remaining areas where fire had not yet been detected by MODIS.
Variable importance revealed strong associations between most pyroregions and the amount (NPP) and timing of plant growth (NDVI; Figure 3b), but there were considerable differences in the importance of other metrics.The monsoon tropical savanna, for example, had a strong positive association with the median area burned and hotspot density (pyroregion b; Figure 3b; Figure S3), whereas the agricultural zones of southern Australia were characterized by fires that predominantly burned only during the day (pyroregion c; Figure 3b; Figure S3).The discreteness of fire seasons varied among pyroregions; for example, the tropical savanna has a discrete fire season coinciding with the onset of the dry season (Figure 4b), whereas fire was observed year-round in the Pilbara (Figure 4f) and Brigalow regions (Figure 4h).See Figure S3 for the relationships between each pyroregion and the fire variables used for the regionalization, and Figure 5 for the major land cover types of the pyroregions.
Pyroregions were strongly associated with the climate.The random forest model discriminated the pyroregions with 84% accuracy based on five climate variables, demonstrating the tight climatic signatures of Australia's pyroregions (Table S1 for class-specific accuracy).Annual precipitation and annual mean temperature were the two most important predictors of pyroregions on average, but again, there was considerable variability in the importance of climate variables among the pyroregions (Figure 3c).For instance, the temperate forest pyroregion (class d) was instead most strongly associated with relatively high precipitation during the driest quarter (Figure 3c).

| Potential for regime shifts and novel fire regimes
Shifts in the projected climates indicated extensive risk of reorganization of Australia's pyroregions.Under the lowest-emission scenario that we analyzed (SP1-2.6),41% of the extent of the pyroregions, on average, were projected to occur outside their historical climate hypervolumes (i.e., mean of proportional change; Figure 6c).Under the higher-emission scenarios, SSP2-4.5 and SSP3-7.0,an average of 65% and 85% of pyroregion extents were projected to occur outside their climate niches (Figure 6c).Temporal pattern of hotspot density envelopes (Figure 6a).This occurs partly because these northern pyroregions occupy narrower climate niches than cooler temperate pyroregions; for example, the tropical monsoon savanna pyroregion has a considerably narrower niche than the wet temperate forest pyroregion (Figure 7).

Continental climate novelty was widespread but less likely
than pyroregion-specific novelty.On average, continental climate novelty occurred over 31% (SSP1-2.6),52% (SSP2-4.5),and 75% (SSP3-7.0) of the pyroregion extents (Figure 6c).Novelty was most concentrated in northern Australia (Figure 6b).Under middle-ofthe-road projections (SSP2-4.5),most of north-western Australia is projected to occupy climate space that has no present-day analogue on the Australian continent, implying the likely emergence of fire regimes that also lack modern counterparts.The concentration of novel climates in tropical regions primarily occurs because these already-warm areas are first to depart present-day climate space, and secondarily occurs because of a slight moderating effect of the southern and eastern coastlines on projected temperature (Figure S8).

| DISCUSS ION
Climate change is already causing some fire regimes to depart from their range of historical variability (Canadell et al., 2021;Hanes et al., 2019), sometimes carrying severe consequences for ecosystems (Kelly et al., 2020) and societies (Filkov et al., 2020).Our data-driven regionalization, based on up to four decades of satellite observations of fire, provides the most comprehensive delineation of Australia's pyroregions to date.Our analysis identified 18 distinct pyroregions.Large areas of most pyroregions are projected to shift outside of their historical climate niches.Smaller but still very substantial portions of Australia's pyroregions are expected to shift into climate space that lacks present-day analogs, implying high uncertainty in predicting future regime characteristics in those locations.
While intrinsic climate-vegetation-fire feedbacks make it difficult or even impossible to accurately predict the nature of regime shifts using correlative statistical approaches (Bowman et al., 2014), our pyroregion niche approach provides a way to identify locations most at risk of fire regime shifts.

| Regionalization of fire regimes
Millions of years of fire in the Earth system have created a quasisteady state arrangement between climate, vegetation, and fire, giving rise to fire regimes with consistent qualities (Bowman et al., 2020).These regimes may already have begun to shift beyond their historical arrangements.Productivity was the overwhelming determinant of the pyroregions, reinforcing the importance of fuel load in determining Australian fire regimes (Murphy et al., 2013).The strong pyroregion associations with annual precipitation (27% variable importance) and mean annual temperature (23%) likely reflects the key limitations on fire in Australia, in which the arid and forested zones are primarily limited by fuel production and dryness, respectively (Boer et al., 2021;Bradstock, 2010).
The 18 Australian pyroregions displayed characteristic signatures.Fire is most common in the monsoon tropical savanna, where the fire season begins and ends in synchrony with the dry season (Figure 4b).The fire season is less defined in wet temperate forests, where infrequent fire is spread over much of the year, with occasional summers having much higher hotspot densities (Figure 4d).
The agricultural zones were characterized by fires in autumn burning crop residues and renewing pasture (Figures 4c,n).Our analysis revealed more complexity than expected in the arid interior and west of Australia, where fire is infrequent and strongly controlled by intra-annual climate modes such as La Niña (Figure 2j; Murphy et al., 2013).This infrequency relative to the length of the satellite record likely led to the relatively high classification uncertainty in this region (Figure 3a).While our analysis revealed 18 pyroregions, all clustering analyses are subject to uncertainty in the optimal number of clusters.For example, Archibald et al. (2013) selected five global pyromes even though BIC continued to marginally improve with additional clusters, while there was no clear choice of complexity in Pais et al.'s (2023) global regionalization.
Because fire regimes are multidimensional abstract concepts, there is no one such variable that encapsulates a fire regime, rather an ensemble of variables are diagnostic of different fire regimes.Furthermore, fire regime classifications are strongly shaped by scale, leading to the coining of pyromes and pyroregions, analogous to biomes and ecoregions, respectively.Early global analyses based on satellite data described 5-8 global pyromes (Archibald et al., 2013;Chuvieco et al., 2008).Since then, studies have used ever-increasing satellite observations to reveal more complex regionalizations at finer spatial scales (e.g., Galizia et al., 2023;Pais et al., 2023;Zubkova et al., 2022).Notably, Zubkova et al. (2022) delineated eight Australian and nine African fire regions using fire-relevant climate variables rather than fire metrics directly.While some of their regions shared similarities with ours (e.g., northern Australia), their results did not distinguish some regimes, such as the small autumn human-engineered fires of pyroregion n from the larger uncontrolled summer wildfires of the adjacent pyroregion p (Figures 3a and 4).
These are fundamentally different fire regimes despite occupying similar climate space, illustrating that a climate-based regionalization can underestimate non-climate drivers of fire and confound cause and effect, as explained by Archibald et al. (2013).
The recent delineation of 15 global pyromes by Pais et al. (2023) also revealed substantial differences to our results.with among the highest fire frequency on earth (2-5 years) with a large arid zone where fire returns at multi-decadal intervals (Murphy et al., 2013).
While regionalization differences are expected given different methodologies and objectives, such differences highlight the pressing need for standardization.The diverse regionalization outcomes seemingly stem from differences in (i) variables used to define fire regimes, (ii) statistical methodologies, (iii) desired spatial complexity, (iv) domain and grain size, and (v) region-specific expertise.While these prior analyses all advance the field of pyrogeography, there is a clear need for the pyrogeography community to develop a consensus approach for defining pyromes, pyroregions and fire regimes that are globally consistent (e.g., as has recently been done for Earth's ecosystems; Keith et al., 2022).Because fire regimes are simplifications of multi-dimensional phenomena, the most critical aspect of standardization would involve a conceptually robust, agreed-upon definition of the minimum important dimensions that define fire at various geographical scales.Crucially, global standardization must ensure adequate representation of regional fire ecology expertise as a prerequisite to arriving at sensible regionalizations, rather than solely relying on remote sensing products.

| Fire regimes in a warming world
Climate change is exacerbating fire weather, landscape flammability, and natural ignition sources from lightning (Abatzoglou et al., 2019;Bowman et al., 2020;Pausas & Keeley, 2021;Romps et al., 2014;Veraverbeke et al., 2017).Its effects on fuel quantity are less clear because climate change is likely to increase the intensity of both droughts and flooding rains in many regions (King et al., 2020;Wasko et al., 2021).Intensifying drought and heat stress will increase plant mortality (Choat et al., 2012;Marchin et al., 2022), leading to a transitory increase but potentially longer-term decrease in dead fuel loads (Ruthrof et al., 2016).A countervailing factor is elevated CO 2 which is predicted to have a fertilization effect on vegetation, but the extent of fertilization is uncertain because of Australia's nutrient-limited soils (Jiang et al., 2020) and the overwhelming effect of increasing intensity and duration of droughts (Brodribb et al., 2020;McDowell et al., 2022;Rifai et al., 2022).There are already indications of fire regime changes, including the collapse of obligate seeder forests in areas where fire became too frequent for seedlings to mature (Bassett et al., 2015;Le Breton et al., 2022;Nolan et al., 2021), and the destruction of an ancient Gondwanan plant community that persisted in fire refugia that has become more conducive to lightning fire (Harris et al., 2018).Because intrinsic feedbacks and alternative stable states pose significant challenges to predicting the nature of regime shifts (Bond et al., 2005;Bowman et al., 2014;Staver et al., 2011a), we instead predicted locations most at risk of regime shifts by identifying the extent to which projected climates occur within the climate niches of Australia's pyroregions.

Vulnerability to regime change was predicted in all Australian
pyroregions, but it was concentrated in tropical and hot-arid pyroregions (i.e., northern Australia).The narrow niche of the monsoon tropical savanna, for example, is projected to cause it to depart its climate niche earlier than southern temperate pyroregions (Figure 7; Figure S7).The narrow niches of the tropical pyroregions parallel similarly narrow climate niches in tropical plant and animal species that are closer to absolute physiological limits making them vulnerable to local extinction (Grinder & Wiens, 2023).Tropical and hot-arid pyroregions were not only at higher risk of departing their own niches, but also at higher risk of drifting into continentally novel climates.This finding echoes other studies that predict equatorial concentrations of novel or extreme climates because these already-warm locations are first to depart present-day climate space (Beaumont et al., 2011;Dahinden et al., 2017;Williams & Jackson, 2007;Xu et al., 2020).The emergence of novel climates implies a lack of analogues (Williams & Jackson, 2007) that can inform our expectations of future fire regime states.Although northern pyroregions were more likely to occupy novel climates, all pyroregions were nevertheless expected to shift substantially outside of their historical niches, suggesting the coming period of flux will involve widespread reorganization of fire regimes across much of the Australian continent.
The paleo record provides evidence that fire regimes were vastly different under past climates.Sedimentary charcoal records, for instance, indicate less fire during the productivity-limited cold intervals and more fire in warm intervals in the Quaternary (Bowman et al., 2009;Power et al., 2008).It is therefore entirely precedented that climate shifts cause widespread changes to fire regimes.The extent of reorganization of modern pyroregions will depend partly on whether the fire regime niches will persist within migrating climate zones and if species are able to track these geographically shifting niches.Long generation times of trees seemingly provide negligible adaptive capacity to rapid environmental change, although phenotypic plasticity may facilitate acclimation to new climates (Brodribb et al., 2020).Hysteresis in climate-plant-fire relationships or changes in human fire management may either help resist or exacerbate state change (Bowman et al., 2022).Such emerging management strategies include reintroducing herbivores to reduce fuels (Johnson et al., 2018), using low-flammability plants as green fire breaks (Cui et al., 2019), mechanically removing woody fuels (Furlaud et al., 2023), letting fires burn under appropriate conditions (Boisramé et al., 2017), reseeding populations that collapsed under frequent fire (Bassett et al., 2015), rapidly deploying fire-fighting teams to protect sensitive biodiversity (Hankin et al., 2023;Kelly et al., 2020), and assisting forests to reach lower-flammability mature states (Zylstra et al., 2022(Zylstra et al., , 2023)).Crucially, fire management will likely benefit from the ancient wisdom of Indigenous-led fire management, involving fine-tuned application of fire when each vegetation type requires it (Bowman & Sharples, 2023;Steffensen, 2020).as of fire modify fire's niche (Balch et al., 2017), but the wisdom of to maintain an ecological system outside of niche requires careful consideration (Nimmo et al., 2022).
While fire activity will certainly change at different spatiotemporal scales (pyromes, pyroregions and fire regimes), a crucial remaining question is the rate, magnitude and direction of such shifts.This may depend in part on whether a regime is limited by productivity (e.g., arid Australia) or dryness (e.g., temperate forests; Bradstock, 2010;Russell-Smith et al., 2007;Williamson et al., 2016).
Climate projections generally predict an atmospheric drying trend (Clarke et al., 2022), implying higher confidence of regime changes in Australia's dryness-limited temperate forests.Due to higher uncertainty in future rainfall (Grose et al., 2020;Ukkola et al., 2020), and the dominant cyclical influence of multiple climate modes on Australia's rainfall (Abram et al., 2021), this implies higher uncertainty of changes to productivity-limited arid regimes.Thus, the key outstanding question-the nature of regime shifts-is elusive because fire regimes depend on vegetation type, which itself depends on both extrinsic climate factors, intrinsic feedbacks, and human factors (Bond et al., 2005;Bowman et al., 2014).The ongoing development of dynamic vegetation models that accurately model fire remains key to resolving this problem (Hantson et al., 2016;Langan et al., 2017).

| CON CLUS IONS
Our data-driven approach to evaluating risk of regime shifts revealed that Australia's 18 pyroregions are all expected to occupy climate space that is, to varying degrees, outside of their historical climate niches.This finding implies that pyroregions are highly susceptible to regime shifts and highlights the fragility of pyroregion distributions in a changing climate.Some pyroregions will occupy climate space that is without analogues on the Australian continent, implying high risk of the emergence of no-analogue regimes.The pyroregions form coherent units from which to evaluate changes to fire regimes, and their boundaries can be aggregated as required, or updated as satellite datasets grow and as regimes change.While future regime states remain highly uncertain, we can be confident that this century will be marked by significant reorganization of fire regimes in Australia and beyond.The extent of fire regime reorganization will ultimately depend on the emissions pathway society takes and the degree to which fire management adapts to the profound challenges of managing fire in a more flammable world.

ACK N O WLE D G E M ENTS
use has been criticized as being less relevant to policy development(Hausfather & Peters, 2020).We assembled the following five climate variables (Pearson's r < |.8|) using historical and projected gridded data from WorldClim v2.1(Fick & Hijmans, 2017): (i) annual precipitation (mm), (ii) precipitation in the driest quarter of the year (mm), (iii) mean temperature of the driest quarter (°C), (iv) annual mean temperature (°C), and (v) temperature seasonality (standard deviation of monthly mean temperatures).Given these variables are long-term averages, they are primarily related to the first two switches of fire (fuel and dryness) , BCC-CSM2-MR(Wu et al., 2019), GISS-E2.1-G(Kelley et al., 2020), MIROC6 (Tatebe et al., 2019), MPI-ESM1.2-HR(Müller et al., 2018).Our selection Proxy of typical fuel loads Mean of annual net primary productivity (NPP), square root transformed (vi) MODIS normalized difference vegetation index 14.NDVI summer-winter difference Spatiotemporal pattern of plant growth and dryness Mean NDVI in August was subtracted from that in February, such that larger values indicate relatively greener vegetation in the Austral summer Note: Variables are grouped by data source.All products were produced at 0.25° resolution.Elements of Australia's fire regimes and their fuels.These 12 measures of fire activity and two measures of fuel were used to regionalize Australia into pyroregions that represent areas with internally consistent fire regimes.For visualization purposes, metrics are shown untransformed (g, k, l), log-transformed (a, c-f, i, j), root-transformed (m), or on their final transformed scales in the case of unitless variables (b, h, n).See Table 2 for a description of these variables.of these GCMs involved screening for a transient climate response score in the range 1.4-2.2°C,as recommended by Hausfather et al. (2022) as a way to avoid the 'hot model' problem in which a subset of CMIP6 models seemingly cause an upward bias to resulting ensemble temperatures.See Figures S4-S6 for maps showing how the individual climate models differed from the ensemble average.
with low interannual variability (Figure 2j); this gave rise to high F I G U R E 3 Australia's pyroregions and their features.(a) The regionalization of Australia's pyroregions and associated assignment uncertainty.Pyroregions are ordered from most to least certain.The fire (b) and climate (c) associations of each pyroregion.Circle size shows the random forest variable importance.Colours show the median value of each variable, demonstrating the direction of the association relative to 0.5 as the mid-point of the data range.In both (b, c), variable importance was scaled to sum to one for each pyroregion.See Figures S3 and S7 for violin plots showing the full distributions of these variables.pyrodiversity(Figure2g), with some cells in this region containing patches that were last burnt in each of the 21 years of the MODIS data analyzed.Variability in burned area was much larger in the arid interior (Figure2j), where individual daily patch areas of up to 542,400 ha were recorded (Figure2c,d).The relative difference between day and night hotspots provided a signature of humancontrolled daytime fires in the agricultural regions of southern Australia (low values, i.e., fires relatively less frequent at night; The geographic patterns of shifts in niches indicate that pyroregions in the tropical and hot-arid north of Australia are most at risk of departing from their historical climate F I G U R E 4 Hotspot signatures of Australia's pyroregions demonstrating fire seasons.Heat maps show the pyroregion-specific patterns of hotspot density (square root scale) over the annual cycle from 2001 to 2021, illustrating patterns in the timing and duration of fire seasons as well as interannual variability.Panel labels correspond to the pyroregions in Figure 3.

F
Proportional land cover of the pyroregions.(a) The major land cover types of Australia, simplified from the Australian Dynamic Land Cover Dataset version 2.1 (McIntyreet al., 2015).(b) The proportional land cover of the pyroregions, with land cover comprising less than 1% of a pyroregion not shown (so some columns do not sum to 1).F I G U R E 6Projected changes to the climate suitability of Australia's pyroregions under three climate change scenarios for 2081-2100.Maps indicate whether the projected climate is novel to the pyroregion (a) and the continent (b).Orange and red show that one or both hypervolume methods, respectively, predict that the future climate will occur outside of the historical hypervolume.Blue indicates that both methods predict that the future climate will occur within a pyroregion's historical climate hypervolume.(c) The proportion of each pyroregion that is projected to occur outside its historical climate hypervolume.The bars show the mean of the loose and tight wrap hypervolume methods, and the dashed line shows the mean of the bars.
For example, their results aggregated into a single pyrome the northern monsoon tropics with arid central Australia, lumping together an area F I G U R E 7 Example comparison of the climate niches of the tropical monsoon savanna and the wet temperate forest pyroregions.The monsoon savanna pyroregion has a much tighter climatic niche than the temperate forest.Solid lines show the historical climate and dotted lines show the projected climate under a middle-of-the-road Shared Socioeconomic Pathway (SSP2-4.5).See Figure S7 for the distributions of the historical climate variables for each pyroregion.
burned area Monthly rasters from 1982 to 2018 Otón et al. (2021) provide a globally consistent monthly grid of burned area at a spatial resolution of 0.25°.Descriptions of metrics used to characterize Australia's fire regimes.
While this product is coarser than MODIS burned area, it provides a longer time series from which to evaluate interannual variability, which is vital for quantifying longer fire return intervals in temperate forests or arid regions Global fire emissions Monthly rasters from 2002 to 2020 van Wees et al. (2022) used the Global Fire Emissions Database framework to estimate greenhouse gas emissions from biomass burning at 500-m spatial resolution MODIS NPP Annual rasters from 2000 to 2020 MODIS net primary productivity (MOD17A3HGF; Running & Zhao, 2019) provides gridded estimates of the annual amount of energy stored as organic matter by plants at a 500-m spatial resolution, which we interpreted as a proxy of fuel production