Identifying and Attributing Regime Shifts in Australian Fire Climates

This paper introduces and analyzes fire climate regimes, steady-state conditions that govern the behavior of fire weather. A simple model representing fire climate was constructed by regressing high-quality regional climate averages against the station-averaged annual Forest Fire Danger Index (FFDI) for Victoria, Australia. Four FFD indices for fire years 1957–2021 were produced for 10 regions. Regions with even coverage of station-averaged total annual FFDI (ΣFFDI) from 1971–2016 exceeded Nash–Sutcliffe efficiencies of 0.84, validating its widespread application. Data were analyzed for shifts in mean, revealing regime shifts that occurred between 1996 and 2003 in the southern states and 2012–2013 in Queensland. ΣFFDI shifted up by ~25% in SE Australia to 8% in the west; by approximately one-third in the SE to 7% in the west for days above high fire danger; by approximately half in the SE to 11% in the west for days above very high, with a greater increase in Tasmania; and by approximately three-quarters in the SE to 9% in the west for days above severe FFDI. Attribution of the causes identified regime shifts in the fire season maximum temperature and a 3 p.m. relative humidity, with changing drought factor and rainfall patterns shaping the results. The 1:10 fire season between Regimes 1 and 2 saw a three to seven times increase with an average of five. For the 1:20 fire season, there was an increase of 2 to 14 times with an average of 8. Similar timing between shifts in the Australian FFDI and the global fire season length suggests that these changes may be global in extent. A trend analysis will substantially underestimate these changes in risk.


Introduction
Over the past two decades, Australia has experienced wildfires of increased frequency and severity due to climate change [1,2]. The 2019-2020 fire season, the black summer, was the worst on record for area burned and property loss [1,2]. Two outstanding issues in understanding the current level of fire risk are: how much of that risk is due to humaninduced climate change, and how much of an increase do we need to plan for?
Both aspects are difficult to quantify. Fire risk indices require high-quality input data that are often unavailable. Existing data are of limited quality, spatial coverage, and continuity. The risk of wildfire also depends on land surface and cover conditions, and in response to increasingly severe fire weather, is also highly nonlinear. Bradstock [3] described four fire switches that dial up fire risk: biomass production, biomass readiness to burn, fire weather, and ignition sources. Climate influences all of these, the third most directly. All four are needed for a comprehensive assessment of fire risk at a given location [3]. The typical fire risk at a place taking account of these characteristics make up the generally accepted definition of a fire regime [4].
This paper focuses on understanding and quantifying the external climatic conditions that contribute to regional fire climate, i.e., the climate entering a region before considering its land use and land cover, level of fuel, dryness, and topography. The incoming regional error introduced by this substitution is small and does not affect the overall results. The regression relationship is described by the following: FFDI = A + B × P + C × T max FS + D × T max 90 + E × C3pm (3) where A, B, C, and D are constants. The conceptual model used in building a fire climatology (pyroclimatology) is based on similar approaches used in hydroclimatology, originally articulated by Budyko [27]. Variables such as rainfall and potential evaporation (Ep) averaged over time incorporate general land surface characteristics via various feedback mechanisms. For example, simple models that utilize changes in mean P and Ep can be used to estimate changes in catchment yield if more detailed models and data are unavailable [5,6]. The relationship between sensible and latent heat on unsaturated land surfaces is fairly linear [28,29], resulting in meaningful correlations between variables such as T max and P. On shorter timescales, windspeed is used to establish short-term evaporative demand. On longer timescales, regional average rainfall, temperature, and radiation can be used, e.g., [30,31].
Fire climates are at the drier end of hydroclimates, offering the potential for heat and atmospheric moisture deficit to be similarly represented. Accordingly, a simple fire climate model need not rely on the variables in Equation (1) but on those that integrate over seasonal timescales. This approach treats fire climate and fire weather separately. Climate governs fire weather through thermodynamic constraints, rather than being the accumulation of fire weather statistics. The variables used to represent fire weather are also affected by quality issues. The daily FFDI requires windspeed and relative humidity, but because of the difficulty in developing homogenous records, Australia still lacks high-quality data sets for each.
Of the variables in Equation (3), T max FS represents atmospheric heat energy, P represents moisture variability within the broader environment, T max 90 represents the drought factor at landscape scale, and C3pm is a proxy for atmospheric moisture. If the model is to be physically representative, constants A, B, C, and D will be generally applicable across regions other than Victoria and point and areal relationships will obey the distributive law, e.g., where averaged station values of the FFDI across a region correspond to estimates generated with spatially averaged inputs. This is routinely carried out for variables, such as temperature and rainfall, but needs to be verified for more complex variables such as fire danger indices.

Model Evaluation
The model was evaluated by comparing the results with an updated record of FFDIs covering 39 stations around Australia from 1971 to early 2017 (LH2019) [15,16]. It contains the lowest, 25th, 50th, 75th, 90th, 95th, and 97th percentiles of the FFDI and the highest estimate for each season (summer, autumn, winter, and spring). It also includes mean values of DF, KDBI, T max , P, Pdays, RH, and V. The fire year median daily FFDI summed annually (MFFDI) and the annual mean daily 97th percentile FFDI (97FFDI) were used in comparisons. The MFFDI is approximately 80% of ΣFFDI depending on the degree of skewness, and both are similar at higher values. The 97FFDI falls between Days Hi+ and Days VHi+ and was compared with Days VHi+ and Days Sev+, except for Tasmania where the latter was not recorded. These data were produced using Lucas' (2009) original method summarized in Equation (1) with minor modifications [15,32]. The LH2019 data were converted into fire year means for each station. All the dates provided in the paper are for fire years unless otherwise indicated, i.e., 1972 refers to the 1972-1973 fire season.
The baseline station data averages for Victoria were spatially unweighted, so we tested the results of Equation (3) for ΣFFDI and Days Sev+ for different station combinations. The best results were from the 7 stations used, closely followed by 6 (omitting Bendigo), followed by the original 9 supplied. Testing the 4 stations used in the LH2019 data set against the state average provided slightly lower estimates, with an adjusted r 2 value of 0.90 against 0.93 for ΣFFDI and 0.69 against 0.72 for Days Sev+. The reference data are Climate 2023, 11, 121 5 of 54 slightly more spatially representative than the data used for evaluation. We also used the Nash-Sutcliffe efficiency measure [33] to compare time series. It produces results similar to the r 2 statistic where time series perform similarly but has greater penalties where they differ. It is commonly used to measure the performance of hydrological models [34]. Hydrologic variables fluctuate widely as do the more recent FFDI estimates so this is a more rigorous measure. Moriasi et al. [34] consider anything below 0.50 as unsatisfactory, 0.50-0.70 as satisfactory, 0.70-0.80 as good, and above 0.80 as very good.
The longer time period for the HL2091 data covers the training period and an additional 8 years (1971,(2010)(2011)(2012)(2013)(2014)(2015)(2016). Station data from the LH2019 data set were averaged for each state and territory and compared those with the HQD FFDI. Anomalies were used to allow for the different measures of the FFDI in each. The results were influenced by the distribution of climate stations across each region.
We also conducted spatial regressions using the 36 station records in the Climate Change in Australia 2015 (CCIA2015) data set [35], which had 35 in common with the 39 from LH2019. CCIA2015 provided average FFDIs and Days Sev+ over a baseline period of 1981-2010 along the inputs to the FFDI from Equation (1) for each station. We regressed the average T max , RH, and V as in Equation (4) for both data sets. This served as an additional check and allowed us to compare spatial and temporal sensitivity of selected inputs to Equations (1) and (3).

Regime Testing and Attribution
Testing for regime shifts and inhomogeneities was conducted using the Maronna-Yohai [22] bivariate test. This test has been widely used to detect inhomogeneities in climate variables [36][37][38][39], decadal regime shifts in climate-related data, and step changes in a wide range of climatic time series [40][41][42][43][44]. The test adapts the formulation of Bücher and Dessens [37] testing a single serially independent variate (x i ) against a reference variate (y i ) using a random time series following Vivès and Jones [41]. The important outputs of the test in a time series of length N are: (1) the T i statistic defined for times i < N, (2) the T i0 value, which is the maximum T i value, (3) i0 , the time associated with T i0 , (4) shift at that time, and (5) p, the probability of zero shift. Note that i0 is the last year prior to the change but here we provide the first year after the change. A full technical description of the test can be found in Jones and Ricketts [45].
Attributing regime shifts requires different methods to those used to attribute trends, which identify signal emergence using the signal-to-noise model. Changes in mean state identified by the bivariate test need to be linked to an underlying cause. Shifts are caused by a mixture of forced and stochastic behavior and can be influenced by feedbacks.
Physically, forced regime shifts are a thermodynamic response to steady-state conditions reaching a critical limit, releasing heat from the ocean into the atmosphere [45,46]. These often coincide with dynamically driven shifts, such as those involved in decadal variability, so are usually diagnosed as such. The difference between a purely dynamic regime shift from one with a thermodynamic component involves a change in the net dissipation of energy rather than a reorganization of existing energy patterns. For example, the codependency between P and T max is due to partitioning between sensible and latent heat-a regime shift with a thermodynamic component will be accompanied by a net increase in sensible heat. Regime changes have previously been attributed to net changes in temperature at the regional to global scale [42,45] and to moisture at the hemispheric and global scale [47]. Both are important contributors to the FFDI.
Under this method, codependent variables are regressed against each other during a period of stationary climate-the period preceding any regime shifts in temperature. The predictor is used to estimate the prediction over the subsequent period to see whether it deviates from observations due to external factors. The main pairs used are T max and P, and minimum temperature (T min ) and T max . If the variables being predicted shift independently Climate 2023, 11, 121 6 of 54 of this relationship under stationarity, the residual will contain the forced component which can be detected as a break-point. Regime shifts on land can be associated with shifts in adjacent sea surface temperatures, especially those in T min . T max is forced if it shifts relative to P. This method was developed and used to determine that the temperature in SE Australia was stationary until 1968, when T min shifted followed by T max in 1972. A forced shift in T max followed in 1997-1998 [42]. Work in progress estimates the stationary period across Australia ended between 1956 and 1972.
Paired analyses using one variable as the reference and testing another for shifts were also conducted. If 2 variables contain the same shift, the bivariate test will not identify a step change. This may be due to 2 shifts responding to a common cause or a cause-and-effect sequence where one forces the other. Other results may indicate partial attribution, nonlinear responses, or independent shifts due to a separate process, including inhomogeneities. In short sequences, climate variability may also be a factor. When the timing coincides with a previously identified regime shift, the size and direction of the shift compared to the original shows the degree of attribution. Paired shifts with the same timing show each regime shift is larger than the size of the interaction between the two. If the timing is very different, then the shifts are likely to be due to separate processes.
The t-test with unequal variances from the Real Statistics package for Excel [48] was also run, producing Cohen's d variable [49]. The d variable measures the standardized difference between two means by pooling the variance, unity being one standard deviation. The result was combined with sample size to calculate the power of the test with respect to the mean change. We also calculated the percentiles of new regime means with respect to the preceding regime's distribution and exceedance of Regime 1's 90th and 95th percentile by Regime 2. This allowed nominal changes in return periods for the 1 in 10 and 1 in 20 fire seasons to be calculated. Data produced for the paper can be found in the Supplementary Information.

Model Construction
This section describes the model construction and results developed using the baseline homogenized FFDI data for Victoria. The model constants and likelihoods are shown in Table A1. Regressions for ΣFFDI, Days Hi+, and Days VHi+ show very good performance. Days Sev+ shows good performance according to the Nash-Sutcliffe efficiency coefficients [33,34]. The FFDI tends to become nonlinear at higher values, hence the lower performance for the latter ( Table 1). The regressions themselves are shown in Table A1. Visually, the correspondence between the reference and regression models is also very close (Figure 1), showing the same progression from ΣFFDI to Days Sev+.  (3)) showing the multiple correlation, adjusted r 2 , and the Nash-Sutcliffe efficiency coefficients followed by standard error in base units and percentage of the average.

Variable
Multiple  The model predictions were then compared with the LH2019 data. These data were cleaned up considerably from earlier versions, particularly regarding relative humidity and windspeed, but as we show later, still contains some inhomogeneities. Four of the original stations were available: Melbourne, Laverton, Mildura, and Sale, which were averaged to compare with the statewide mean. As mentioned in the methods section, the coverage was not quite as good as the seven-station reference data.
The pairs of MFFDI and ΣFFDI, and 97FFDI and Days VHi+ were compared for the years 1971-1972 and 2016-2017 using the Nash-Sutcliffe efficiency coefficient ( Table 2). The periods 2011-2012 to 2016-2017, were also compared separately as these were not used in the model regression. In both cases, the performance of all variables is higher for the latter period. The results for P and Tmax are also shown and reveal that the more recent records are more reliable. The resulting FFDI time series are also shown in Figure 2. The residuals were investigated to identify any missing factors but proved to be random.  The model predictions were then compared with the LH2019 data. These data were cleaned up considerably from earlier versions, particularly regarding relative humidity and windspeed, but as we show later, still contains some inhomogeneities. Four of the original stations were available: Melbourne, Laverton, Mildura, and Sale, which were averaged to compare with the statewide mean. As mentioned in the methods section, the coverage was not quite as good as the seven-station reference data.
The pairs of MFFDI and ΣFFDI, and 97FFDI and Days VHi+ were compared for the years 1971-1972 and 2016-2017 using the Nash-Sutcliffe efficiency coefficient ( Table 2). The periods 2011-2012 to 2016-2017, were also compared separately as these were not used in the model regression. In both cases, the performance of all variables is higher for the latter period. The results for P and T max are also shown and reveal that the more recent records are more reliable. The resulting FFDI time series are also shown in Figure 2. The residuals were investigated to identify any missing factors but proved to be random.

Regime Detection
The close correspondence between the reference station average, regression model, and LH2019 data provides access to fire climate and input data spanning from 1957-58 to 2021-2022. These time series were analyzed using the bivariate test to detect regime shifts and any inhomogeneities. The timing of breakpoints is often sensitive to the beginning and start date of a time series, so having longer a time series is an advantage. The test can also be sensitive to the timing of adjacent high and low anomalies depending on the direction of any shifts.
The bivariate test results are shown in Table 3. The reference data shifted in 2002-2003 and are the largest of all the estimates. The predicted FFDI for the same period shows shifts at the same time, but only 82-89% of the magnitude the reference data, so the regression model captures the same timing but is slightly underestimated. The blended data sets are made up of the predictions with observations infilled where available, and their shifts are similar to the predicted model over the same period. Over the longer record, the shift dates move earlier to the 1996 and 1997 fire years. These latter differences can be attributed to variations in climate over time, affecting regime means. The period 1972-1975 at the start of the reference period was wetter than the long-term average and 2009-2010 marked the Black Saturday peak. This was followed by a series of very wet and very dry fire seasons, the series ending in 2021-2022 with two successive cool events.

Regime Detection
The close correspondence between the reference station average, regression model, and LH2019 data provides access to fire climate and input data spanning from 1957-58 to 2021-2022. These time series were analyzed using the bivariate test to detect regime shifts and any inhomogeneities. The timing of breakpoints is often sensitive to the beginning and start date of a time series, so having longer a time series is an advantage. The test can also be sensitive to the timing of adjacent high and low anomalies depending on the direction of any shifts.
The bivariate test results are shown in Table 3. The reference data shifted in 2002-2003 and are the largest of all the estimates. The predicted FFDI for the same period shows shifts at the same time, but only 82-89% of the magnitude the reference data, so the regression model captures the same timing but is slightly underestimated. The blended data sets are made up of the predictions with observations infilled where available, and their shifts are similar to the predicted model over the same period. Over the longer record, the shift dates move earlier to the 1996 and 1997 fire years. These latter differences can be attributed to variations in climate over time, affecting regime means. The period 1972-1975 at the start of the reference period was wetter than the long-term average and 2009-2010 marked the Black Saturday peak. This was followed by a series of very wet and very dry fire seasons, the series ending in 2021-2022 with two successive cool events.
The predicted results from 1957-1958 to 2021-2022 are shown with breakpoints and internal trends in Figure 3. All trends are slightly negative and have p values > 0.4, showing that the data form distinct regimes. The arithmetic differences between regime means were 25% for ΣFFDI and 30% for Days Hi+ in 1995-1996 and 54% for Days VHi and 77% for Days Sev+ in 1996-1997.  The predicted results from 1957-1958 to 2021-2022 are shown with breakpoints and internal trends in Figure 3. All trends are slightly negative and have p values > 0.4, showing that the data form distinct regimes. The arithmetic differences between regime means were 25% for ΣFFDI and 30% for Days Hi+ in 1995-1996 and 54% for Days VHi and 77% for Days Sev+ in 1996-1997.

Sensitivity Analysis
To better understand what is driving these changes, we tested the sensitivity of FFDIs to each input variable and analyzed the results for shifts. Two tests were run. The first omitted each variable from Equation (3) in turn and the second increased each in turn by 10%. For each variable we looked at changes in the mean, standard deviation, and the magnitude and timing of each regime shift. The detailed results are shown in Tables A2  and A3. We also tested the input variables for shifts over the period 1957-2021 (Table 4).

Sensitivity Analysis
To better understand what is driving these changes, we tested the sensitivity of FFDIs to each input variable and analyzed the results for shifts. Two tests were run. The first omitted each variable from Equation (3) in turn and the second increased each in turn by 10%. For each variable we looked at changes in the mean, standard deviation, and the magnitude and timing of each regime shift. The detailed results are shown in Tables A2 and A3. We also tested the input variables for shifts over the period 1957-2021 (Table 4). These results show that P and T max 90 contribute to the variability of the FFDI and timing of regime shifts, whereas T max FS and C 3pm affect the magnitude. The change of −2.5% in C 3pm in Table 4 does not qualify as a regime shift, but based on the sensitivity testing, would have had an effect at levels of higher fire danger.

Regional Comparisons
The regression model optimized for Victoria was applied to other states and regions. If the model has a physical basis, it will not need to be modified for different locations. Figure 4 compares the ΣFFDI with median daily ΣFFDIs averaged for each region and Figure 5 does the same for Days VHi+ with the 97th percentile FFDI. The former provides a better fit, but some regions fare better than others. This can be explained by the distribution of stations in the LH2019 data with respect to regional average climate. Table 5 shows the Nash-Sutcliffe efficiency between the LH2019 and HQD results for each region. These results depend on whether station coverage in the LH2019 data set is evenly distributed and whether those data are homogenous. Performance for Victoria, New South Wales, and SE Australia is good to very good. Queensland and Northern Territory results are very good for ΣFFDIs and satisfactory for Days Sev+, perhaps because arid to tropical climates are included in the same region. South Australia is satisfactory; the four stations represented are in the cooler and wetter southeast, whereas much of the state is a low-rainfall, hot climate. Tasmania is unsatisfactory due to the bias of stations towards the drier and warmer eastern side of the state. Western Australia has a distribution biased towards the coast and widely varying climates. SW Western Australia has three records that are badly affected by inhomogeneities. For the latter, T max N-S efficiency is very good, so the problems are with the other variables.

Regional Comparisons
The regression model optimized for Victoria was applied to other states and regions. If the model has a physical basis, it will not need to be modified for different locations. Figure 4 compares the ΣFFDI with median daily ΣFFDIs averaged for each region and Figure 5 does the same for Days VHi+ with the 97th percentile FFDI. The former provides a better fit, but some regions fare better than others. This can be explained by the distribution of stations in the LH2019 data with respect to regional average climate.         Table 5 shows the Nash-Sutcliffe efficiency between the LH2019 and HQD results for each region. These results depend on whether station coverage in the LH2019 data set is evenly distributed and whether those data are homogenous. Performance for Victoria, New South Wales, and SE Australia is good to very good. Queensland and Northern Territory results are very good for ΣFFDIs and satisfactory for Days Sev+, perhaps because arid to tropical climates are included in the same region. South Australia is satisfactory; the four stations represented are in the cooler and wetter southeast, whereas much of the    Table 5. This is despite a slightly warmer and drier climate represented in the state averages (+1.1 • C and 61% P). There is a subtle change in balance between the two over the period of record where the MFFDI increases faster than the ΣFFDI.
For South Australia the outcome was poorer, with r 2 values of 0.69, 0.66, and 0.60 due to two arid zones and two temperate stations, representing a state with a very large arid zone. The station and state averages for T max and P differ widely, and the correlations are better for T max than P.
Tasmania has the reverse issue, with the two stations (Launceston, Hobart) being warmer and drier than the state average. R 2 values are 0.69, 0.66, and 0.60. Windspeed registered a shift of 17% in 1993, possibly influencing the balance between the two data sets.
Southeastern Australia covers NSW to just north of Sydney, SA, just west of the Eyre Peninsula, Victoria, and Tasmania. This region contains 12 stations with an even geographic spread, so is more representative of the regional average climate. T max and P data for the 1971-2016 period are 20.6 • C and 21.0 • C and 638 and 627 mm, respectively. The MFFDI/ΣFFDI r 2 is 0.88 and 97FFDI/Days VHi+ and Days Sev+ are both 0.86. These are similar to the results for Victoria, showing the impact of good regional coverage.
Queensland shows r 2 values of 0.87, 0.67, and 0.64. The former is probably due to more comprehensive state coverage of eight stations, although with a coastal bias. R2 values for T max and P are 0.87 and 0.92 despite differences in T max (28.3 • C and 30.3 • C) and P (1053 mm and 647 mm, stations and state, respectively).
Northern Territory shows r 2 values of 0.85, 0.76, and 0.69, probably due to the three evenly spaced stations even though the fire season in Darwin is distinctly different to that in Tennant Creek and Alice Springs. This also demonstrates the capacity for Equation (3) to represent different climatic patterns.
Western Australia has low r 2 values of 0.60, 0.50, and 0.48. WA contains several climates which may be unevenly represented (as is the case for SWWA); however, inhomogeneities may also be affecting the results, including an 8% shift in windspeed in 1994.
For SWWA, the individual stations do not represent the regional average very well, with r 2 values of 0.37, 0.23, and 0.20 for the three comparisons. Of the three stations, Perth represents regional temperature and rainfall better than Albany or Esperance. Most of the differences are due to rainfall patterns; Albany and Esperance are not very representative of the regional pattern.
Nash-Sutcliffe efficiencies for Southern Australia are also shown in Table 5. The corresponding r 2 values are 0.79, 0.76, and 0.74. They show that when individual stations are widely distributed, the results match those derived from regional mean climate as good to very good for Nash-Sutcliffe efficiency.
All states have inhomogeneities in windspeed and some stations in relative humidity; however, where station coverage is evenly distributed and broadly representative of average climate, results are good to very good for all measures of the FFDI. For those regions where station distribution shows some spatial bias and/or is affected by inhomogeneities, the use of regionally averaged, high-quality input data is likely to provide more realistic results for that region.

Regime Detection
Regime detection was carried out for the four predicted measures of FFDI for the period 1957-1958 to 2021-2022. The results are shown for the ΣFFDI in Figure 6 and Days Sev+ in Figure 7. The results are listed in Table 6. Regime detection was carried out for the four predicted measures of FFDI for the period 1957-1958 to 2021-2022. The results are shown for the ΣFFDI in Figure 6 and Days Sev+ in Figure 7. The results are listed in Table 6.      Queensland shifted in 2012 and the Northern Territory in 2012 and 2017. The latter is provisional because p values are roughly one in four, but the magnitude of change is roughly one standard deviation, which will register if it is sustained over time.

•
These changes show a strong relationship with latitude, where the first changes are south of the sub-tropical ridge moving further north over time. This suggests a strengthening of the Hadley Cell and tropical expansion.

•
Regime shifts in the 2021-2022 fire season were generally lower than those calculated in the Black Summer season in 2019-2020 showing the effect of two wetter and cooler years. Variability in Regime 2 is also much greater than in Regime 1, showing a more intense hydrological cycle, but the regime itself has remained stable. Given the slight underestimation of regime shifts in the model and these cooler conditions, these shifts should be considered as minimum estimates.

Spatial Variations
The LH2019 data set has 39 stations and the CCIA2015 data set has 36 stations, with 35 in common. Common to each station were baseline (1981-2010) inputs into Equation (1) for average annual T max , RH, and V. We conducted a multiple regression of these variables (as described in Equation (4) for MFFDI and 97FFDI) from the LH2019 data and ΣFFDI and Days Sev+ from the CCIA2015 data ( Table 7). The results show a similar pattern of r 2 values to the time series regressions. RH and T max are codependent and largely compensate for each other with RH being the dominant partner. Table 7 also shows the spatial correlations between T max , RH, and V for each of the four indices. Correlations decrease with severity, with RH having the most influence and V the least. The correlation between the V index and the FFDI is negative, showing that less windy locations have higher fire danger. The largest outliers in the predicted results are coastal stations. This is consistent with many coastal locations with high average windspeeds having lower mean FFDIs. Coastal climatologies have more spatially varying wind fields, which can also affect T max and RH depending on the prevailing influences. For 97FFDI and Days Sev+, the influence of windspeed falls below p = 0.05. This is the opposite to the effect of windspeed on fire weather where higher windspeeds are associated with higher fire danger. The LH2019 data show that before instrumentation in 1993, correlations were lower than later on, even reversing in a few cases, a sign of significant inhomogeneities. For RH and T max , spatial and temporal correlations are of the same sign and are more consistent over the period of record.

Attribution
This section explores which variables in Equations (1) and (3) can be classified as the main drivers and shapers of regime changes and those with limited effect. Drivers are involved in forcing the nonlinear response (change in mean) and shapers are those that influence variability and the timing of shifts. Some variables may be both drivers and shapers.
In Section 3.1.3, sensitivity analysis of the factors in Equation (3) showed that T max FS and C 3pm contributed mainly to mean change. P was the largest contributor to shift size, mainly through its contribution to changing variability. T max 90 had a smaller effect on variability and shift size later in the record.
Based on the constants in Equation (1), a 1 m s −1 change in V would change the ΣFFDI by 8.5, a 1 • C change in T max would change it by 12.3, a one-point change in RH would change it by 12.6, and a one-point change in DF would change it by approximately 351. Taking a simple difference for these inputs for Victoria from the LH2019 data from 1971-1995 and 1996-2016 estimates the shift in the ΣFFDI as 351 compared to the measured shift of 645. The dynamic relationship between temperature and moisture on very hot and dry days, amplified by the drought factor, is missing. This amplification shows in the disproportionate increase in days of high fire danger compared to mean FFDI. Equation (1) better represents fire weather, whereas Equation (3) better represents fire climate. This is reflected in the use of T max FS for calculating fire season climate, whereas T max is used for fire weather calculated daily.
Regime shifts in Australian temperatures occurred in the period 1969-1972 generally, from 1978-1979 in northern Australia, and from 1996-1998 mainly in southern Australia. More recently, sea surface temperature shifted in 2009-2010 along the northwest to southern coasts and in 2015-2016 along the northeast to east coast. These influenced shifts on adjacent land. Recent shifts can be difficult to diagnose with confidence due to the shortness of record and variations in climate, whereas older shifts have been detectable for quite some time.
We tested the period from 1957-1958 to 2021-2022 in the HQD and from 1971-1972 to 2016-2017 in the LH2019 data. Most variables are limited to these time periods, but temperature and rainfall-related variables are available from 1910 and 1900, so the longterm record may give different dates and are shown where possible. The most relevant results from the LH2019 data are shown in Table 8. Shifts in all variables are shown in Tables A4-A12.
The two drought factor variables (DF and KBDI), rainfall and rain days, show limited shifts and no distinct patterns. Shifts in rainfall and drought factor in Victoria and SE Australia are associated with the widespread regime changes in 1996-1997, documented in Jones [42]. Increased rainfall in 1973-1974 in the NT (and occurred widely over northern Australia) and 1994-1995 in WA (mainly in the northwest) did not reduce the FFDI, nor did decreased rainfall in SWWA in the late 1960s increase it. This supports the sensitivity analysis which shows that the influence of rainfall is largely interannual. For Equation (3), P and C 3pm showed no patterns that could be related to other climate variables or to regime shifts in the FFDI, except for Victoria.
Shifts were also detected in most records of RH and V. Despite adjustments being made to these records over time, some contain inhomogeneities that can be identified due to changes in instrumentation (windspeed) and anomalous changes in particular stations (RH and V). Jakob [21] described the difficulties in extracting a reliable record from Australian windspeed observations. Troccoli et al. [50] analyzed measurements at 2 m and 10 m heights, subjecting records to rigorous quality control. Windspeeds from 1975-2006 showed a slight decrease at the 2 m height, also noted by McVicar et al. [51], and an increase at the 10 m height. As the two show similar seasonal patterns of change, Troccoli et al. [50] concluded that the 2 m readings were affected by surface modification. Data from seven 10 m stations from 1948-2006 showed little change, so they concluded there were no evident long-term effects in circulation [50]. Recent adjustments to Australian 2 m wind data show an overall decline over time in daily peak wind gusts and reductions in trends in both directions [52]. The disjunct between spatial and temporal relationships between V and the FFDI reported above shows that windspeed is important to fire weather but not fire climates. Accordingly, windspeed is not considered as a contributor to regime changes. RH from the LH2019 data set shifted simultaneously with T max FS in SWWA (1993-1994) and NT (1979NT ( -1980 but in other regions preceded T max FS (shifting 2001-2002in NSW and WA and 1996-1997 in the other regions). The latter timing coincides with the shift to dry atmospheric conditions at the beginning of the Millennium drought in SE Australia as documented by the Bureau of Meteorology [53]. In Jones and Ricketts [47], we identified large-scale changes in RH in the HadISDH data set [54]. shift to dry atmospheric conditions at the beginning of the Millennium drought in SE Australia as documented by the Bureau of Meteorology [53]. In Jones and Ricketts [47], we identified large-scale changes in RH in the HadISDH data set [54].  Another way to conduct nonlinear attribution with the bivariate test is via paired analyses by using one variable as the reference and testing another against it. If the pair contain the same shift, the test will show no change. For example, pairing the different measures of the FFDI for a single region show no regime changes (p > 0.10). This type of change may denote a common origin for both variables being tested or direct cause-andeffect where one forces the other (i.e., full attribution). Other results can show partial attribution, nonlinear responses, or shifts due to an independent process not directly related to change in fire climates, including inhomogeneities. Climate variability may also be a factor in short sequences. Another way to conduct nonlinear attribution with the bivariate test is via paired analyses by using one variable as the reference and testing another against it. If the pair contain the same shift, the test will show no change. For example, pairing the different measures of the FFDI for a single region show no regime changes (p > 0.10). This type of change may denote a common origin for both variables being tested or direct causeand-effect where one forces the other (i.e., full attribution). Other results can show partial attribution, nonlinear responses, or shifts due to an independent process not directly related to change in fire climates, including inhomogeneities. Climate variability may also be a factor in short sequences.
Appendices B.1-B.9 show paired bivariate tests with the ΣFFDI and all input variables from Equation (3), except T max 90 in the top part, with the ΣFFDI and Days Sev+ for the Equation (1) input variables from the LH2019 data set in the bottom part. These results are summarized in Table 9. They must be interpreted with care because of the relatively short and differing record lengths between the HQD and LH2019 records (65 and 46 years, respectively). They show similar influences as the other tests conducted but the results are more complicated. For example, the tests pick up relative changes between P and T max that are climate-driven but did not affect fire regimes. Shifts in 1975-1976 in the LH2019 data set were due a succession of wet years near the beginning of the record. Those in 1979Those in -1980 in northern Australia were the regional expression of a global shift.  In all cases, shifts in the FFDI were consistent with those in V, but the reverse shifts showed almost no change in V from the standard shift against random data. This is interpreted as being due the observer-instrument change-over just preceding the regime shifts in the FFDI.
Variables with minor contributions to regime shifts are C 3pm, T max 90 (not tested in pairs), V, and PDays. P has a moderate influence, mainly in the mid-latitudes. The strongest influences are T max FS, RH, and DF/KBDI. KBDI and PDays are inputs into DF, so DF is the key variable. In Vic, NSW, SA, Tas, SEA, Qld, and WA, regime shifts all occur with a combination of RH and T max FS. For NT, there is too little information, and any shifts should be considered preliminary. The mean of MFFDI and 97FFDI from four stations in central Australia show shifts in 1979-1980 of 26% and 22%, respectively, at p < 0.01, which is consistent with the changes from RH and T max FS in Table 8. For SWWA, applying the same rules would suggest a regime shift in either 1993-1994 if the long-term data from 1910 were applied to T max or 2006-2007 for the 1957-2021 data based on a shift in T max FS, given RH shifted in 1993-1994. On the hunch that adjacent sea surface temperatures (SST) may be playing a role, we subtracted the influence of P and SST from T max FS, which then showed a shift in 2002-2003. This coincided with the shift in Days Sev+ two years following the shift in the other FFDI variables. This suggests that changes in SST influenced the timing of changes there.
With RH as the reference, in most cases, the FFDI showed no shift. However, in reverse, RH underwent a partial shift against the FFDI in most regions. The reductions in RH did not fully convert into a change in the FFDI. In comparison, no shift registered with V as the reference and the FFDI as the test variable, but in reverse, V was almost completely unaffected. The timing of the two shifts was therefore coincidental, with anemometers being introduced a few years before the FFDI shifted. T max FS also showed no shift with respect to the FFDI or vice versa, indicating a substantial contribution to the shift in the FFDI. It did show a number of relative shifts with different timing, some due to climate variability or coinciding with other known shifts. During the period 2010-2013, T max and P shifted relative to each other and the FFDI in some regions. This denotes a warmer period interspersed by very wet years that mask some of the warming. In some regions, this change was seen slightly earlier. The presence of hot and dry years with occasional very wet and cool years later in the record contributes to greater variability in the FFDI.
The most prominent influencing variables are RH and T max FS along with DF. Changes in the latter are small but consistent with the shifts in the FFDI across most fire-prone regions. We propose that fire climate shifts occur in response to shifts in RH and T max FS, where RH reductions feedback onto T max FS by drying out the landscape, which also influences DF through warmer conditions and greater evaporative demand. This suggests that neither T max FS nor RH can force a complete regime shift on their own but can through their combined feedbacks. This is represented in the FFDI algorithm as DF but is probably more closely aligned to regional soil moisture. Work on substituting soil moisture and related factors into the FFDI to forecast fire risk has identified soil moisture as being the most reliable replacement for DF [55][56][57].
Although Equations (1) and (3) detect the same regime shifts in the FFDI (allowing for different lengths of the input data sets), Equation (1) represents fire weather and Equation (3) represents fire climates measured through proxy variables. Based on these analyses, RH and T max FS are the main drivers, with DF/regional soil moisture and P as the main shapers. Equation (3) represents these well enough to produce regime changes but underestimates those measured from daily FFDIs by approximately 15%. The development of high-quality moisture variables free of inhomogeneities, potentially in a regional soil moisture index, may produce more accurate estimates.
Statistical attribution was also undertaken using the t-test at shift points identified by the bivariate test. Most applied different variances based on F-test results, with two records from Tasmania, SW Western Australia, and the Northern Territory using constant variance. The likelihoods of the t-test for the null hypothesis were much smaller than the bivariate test with none being p > 0.05, so the bivariate test is more stringent. Cohen's standardized difference ranged between 0.8 and 1.15 for most shifts, one being a standard deviation. Cohen's d was also used to estimate power and the ratio of making a correct conclusion with respect to the test being in error. The results are included in the Supplementary Data. Of the 10 regions with four indices each (three for Tasmania), 21% were >0.99, 28% were 0.95−0.99, 21% were 0.90−0.95, 21% were 0.80−0.90, and 10% were below 0.80, the lowest being 0.65 with a power of 2:1. Therefore, both the bivariate and t-test results clearly show that our identification of regime shifts is not due to statistical error but is measuring a genuine change in state.
Nonlinear attribution was also carried out to measure the change in fire seasons in Regime 2 with respect to the prior distribution in Regime 1. We estimated the Regime 2 mean as a Regime 1 percentile. Regime 2 means range from the 0.79 percentile for Tasmania and 0.80 for the Northern Territory of Regime 1 and between 0.94 and 0.97 for Queensland (Table 10). The average is 0.90 for all states and the Northern Territory. We also assessed the exceedance by Regime 2 of the 0.90 and 0.95 percentile from Regime 1 for each region and index (the 1:10 and 1:20 fire seasons). These are shown as factors measuring the rate of increase in Table 10. The 1 in 10 fire season of Regime 1 is exceeded three to seven times as often in Regime 2 with an average factor of five. The 1 in 20 fire season is exceeded from twice to 14 times more often, averaging 8 times more often. The Supplementary Data also show the upper limit of Regime 1 is exceeded in 25% and 50% of fire years in Regime 2 depending on location. Not considering local factors, the fire climate model measures the impact of changing external climate on fire danger, so the changes in Table 10 can be considered as a direct response to external forcing. These changes are much larger than anticipated, coming earlier than previous projections. In some cases, they are equivalent to the upper limit for the 2030 changes from CCIA2015 [58]. Any changes in fire risk due to local conditions will be additional to the changes described here.

Discussion
The initial aim of this work was to determine whether publicly available high-quality data free of inhomogeneities could be used to represent fire climates by substituting the more problematic variables used to represent Macarthur's FFDI, especially relative humidity and windspeed. Given its success, the following aim was to detect changes in fire climate regimes across Australia and then to attribute those changes to specific climate drivers.
The result is a regression model using high-quality inputs that can be applied across the widely varying climates of Australia. Regime shifts have been identified across most of those regions, potentially all, if the signs of a recent shift in the NT are sustained. This work has also identified the major drivers of regime shifts to be T max FS and 3 p.m. RH, with fire season P and a drought factor/regional soil moisture shaping those changes; P via interannual variability; and DF by amplifying shifts in T max and RH, both of which also interact with P.
This work also introduces the concept of a fire climate as a steady-state regime that governs the mean and distribution of fire weather, instead of the more common view of climate being the statistical aggregation of weather over a nominal time period. The presence of steady-state regimes in simple and complex measures, such as the FFDI, means that climatology can focus on physical states rather than statistical states. This will affect how historical climate is analyzed, how future climate is projected, any detection and attribution carried out, and how changing climate risk is characterized.

Caveats, Strengths, and Limitations
As the first study of its kind, this work has been exploratory and opportunistic, using data that were at hand or readily obtainable. The baseline FFDI from 1972-2009 was derived from a seven-station record from Victoria and was adjusted for inhomogeneities, mainly in windspeed. Initial analyses of regime shifts using this data were published in Jones et al. [40] along with an analysis of the economic implications. The seven-station average was not spatially weighted, so is best considered as annual anomalies rather than total FFDIs. However, the close correspondence between the six-and seven-station average and the HQD model for the same period (adj r 2 of 0.93) shows these anomalies closely represent the state average anomaly. Predicted shift size of the HQD model compared to the baseline FFDI was underestimated by 11% for Days Sev+ to 19% for Days Hi+, so the model is slightly more conservative than estimates accumulated from daily calculations. This suggests the model is conservative and has room for improvement. This paper updates a report that analyzed shifts ending in the Black Summer fire year of 2019-2020 [58]. That season was a high point in the FFDI for much of Australia but was followed by three La Niña years of lowered FFDIs, two of which are included in this analysis. The additional data have reduced the size of regime shifts in eastern Australia, especially over NSW, slightly increasing p values against the hypothesis of no shift. For the current year, another mild season in most regions has yet to be added. We are confident that the bivariate results associated with null probabilities of~p < 0.10 are conservative, particularly as the t-test results, including those for the NT, were p < 0.05.
The addition of two extra years' climate data has solidified shift dates from the earlier report, some of which were ambiguous (e.g., switching between 2002-2003 and 2012-2013 in NSW due to the wet years in 2010-2011). Adjustments made to the highquality input data undertaken by the BoM have also helped to improve confidence in shift timing and amount, slightly reducing them in some regions. For example, the addition of more remote uplands data in Tasmania has lowered state average temperatures slightly while increasing rainfall. Such adjustments generally improve bivariate test results irrespective of their direction. The first author has been using the bivariate test on the BoM's high-quality data since it first became available-ongoing quality adjustments have gradually made historical shifts more distinct. This is why high-quality proxies are preferred to the direct inputs to Equation (1) if those inputs are of lesser quality.
The reduced shift size using the variables from Equation (3) is probably due to the lack of a direct moisture variable: C 3pm and P are indirect and, combined with T max 90, act as a proxy soil moisture variable. Having access to high quality regional soil moisture indices free of inhomogeneities may remedy this issue. However, three factors show the regime shifts themselves are robust: firstly, they are produced using different combinations of variables (Equations (1) and (3)); secondly, where individual variables were removed in sensitivity testing, the shifts remained; and thirdly, regime shifts involve changes in both sensible and latent heat consistent with our findings on global and hemispheric changes in atmospheric moisture.

Comparisons with Other Studies
Many studies explicitly refer to the nonlinear nature of recent changes in the FFDI record [14,15,32,[59][60][61][62][63]. However, given the lack of an accepted explanation for how such changes could be a response to forcing, most previous studies have analyzed changes as trends, the exception being Jones et al. [40].
Lucas et al. [14], who recorded a change in trend of 10-40% between 2001-2007 compared to 1980-2000 across various sites in their study, suggested a role for decadal climate variability where decadal cycles were enhanced by anthropogenic change. They also noted that some of the events between 2000 to 2006-2007 exceeded the conditions projected for 2050 presented in the same report. Clarke et al. [32], analyzing the high-quality data set developed by Lucas [17] that extended to 2009-2010, noted the 'jump' around 2000 and speculated whether it was due to decadal variability combined with climate change. Applying trend analysis, the 90th percentile daily FFDI reached the p = 0.05 level in approximately half of the records examined. Clarke et al. [32] also noted that the recent changes rival future projections. Sharples et al. [63] also noted this nonlinear increase, which included an increase in extreme bushfires and pyroconvective events.
Sanabria et al. [64] constructed a national spatial climatology from the 78 stations developed by Lucas [17] by interpolating the input data then calculating the FFDI. They used an additional 35 secondary variables to fine-tune the drought factor to account for fuel dryness. They tested the surface with the larger set of stations then narrowed it down to the 38 of high quality used by Clarke et al. [32]. This provided a smoother surface when interpolated and lower FFDI values, so they endorsed the better-quality but sparser network. They investigated return periods of up to 500 years using different extreme event formulations, producing estimates of different 50-year return periods [64]. However, as seen in Section 3.2.4, nonstationarity in the data will skew the results.
Williamson et al. [65] built a national data set of FFDIs from the SILO database of daily data [66] covering from 1900-2011 using the Noble et al. [13] formulation of FFDIs. They combined this with a satellite record of fire activity from 2000-2011 (hotspot detection) to better understand the climatic influence on regional fire activity over 61 regions derived from a set of climate indices. Comparing fire behavior with the ENSO and Indian Ocean Dipole indices identified strong interannual influences on fire behavior, particularly in the south. There were clear relationships between fire activity and FFDIs across different zones, and fire was associated with different FFDI thresholds across these zones. They did not investigate changes over time, relying on the same conclusion as Jolly et al. [67] which determines that the high interannual variability present would obscure any underlying trends.
Dowdy [60] constructed a gridded record of the FFDI from data contained in the Australian Water Availability Project database with wind taken from an NCEP-NCAR reanalysis extending from 1950 to 2016. The length of time recorded allowed differencing between blocks of two to three decades each without resorting to trend analysis. When divided into halves (1951-1983 and 1984-2016), increases in the mean FFDI at the p < 0.05 level were detected in the south, mainly in the SA region, with decreases in the north and northwest [60]. Separating the last half period into quarters, change was dominated by more widespread increases across southern Australia than earlier, with some northern areas experiencing decreases and other increases. Dowdy [60] also pointed out the nonlinear increases in the FFDI after 1999, especially in southern Australia.
Dowdy [60] presented rarer events such as 1-, 5-and 10-year return periods, stating that a longer time series could produce "climatological analysis with minimal influence from natural variability". However, he also referred to the nonstationary nature of the time series, saying that after 2000 in some SE Australian locations during spring, extreme values were similar to those formerly seen in summer. These conclusions are inconsistent as shown by the clear differences in return periods between Regimes 1 and 2 in Table 10.
Harris and Lucas [15] analyzed their national high-quality FFDI station data set up to 2017. As found by Dowdy [60], they detected larger increases in the south, especially in spring. They recorded trends meeting the p < 0.05 level in the 90th percentile FFDI in 44% of the all-station network (17/39) and in the all-station average [15]. They assessed the influence of the Southern Annular Mode, Indian Ocean Dipole, and El Niño-Southern Oscillation on detrended data, finding that all three had strong seasonal and regional influences on interannual variability, separately and in combination, which is useful for seasonal forecasting. They also assessed whether these variables could contribute to the long-term linear trend or whether Interdecadal Pacific Oscillation could contribute to decadal variability, finding neither could explain the observed changes. Anthropogenic change was identified as the cause of the upward trend without the authors being able to pinpoint the specific mechanism involved [15]. They concluded that observed trends have generally been in excess of projections from climate modelling studies. However, our results show that no trend is present. Figures 6 and 7 show distinct regimes.
In Figure 8, Australian and SH RH over land show regime shifts one year apart, so the question arises as to whether the regime changes in fire danger are more widespread. When global fire season length, as discussed by Jolly et al. [67], is compared with the Australian MFFDI by Lucas and Harris [16], both time series shifted in 2002 (p < 0.01 for Australia, p < 0.05 global; Figure 9). The global fire season length is constructed from the number of days when fire danger exceeded its median value from 1979-2013, measured from three different fire danger indices used in the US, Canada, and Australia, and applied to an ensemble of three subdaily reanalysis data sets, as described by Jolly et al. [67]. Although that analysis shared a period in common for Australian average FFDI from Lucas and Harris [16], Jolly et al. [67] found a trend p < 0.05 for all non-Antarctic continents except Australia/New Zealand. The correlation between both time series is 0.58 (p < 0.01), and with regimes removed, considering interannual variability only, it is 0.35 (p = 0.03). Correlation in Regime 1 is 0.28 (p = 0.19) and in Regime 2 is 0.53 (p = 0.07). This is consistent with regional influences becoming less prominent in favor of global influence under forcing. Further work is required to see whether other regions undergo similar shifts but given the scale of changes in T max and RH, similar changes could be expected.
except Australia/New Zealand. The correlation between both time series is 0.58 (p < 0.01), and with regimes removed, considering interannual variability only, it is 0.35 (p = 0.03). Correlation in Regime 1 is 0.28 (p = 0.19) and in Regime 2 is 0.53 (p = 0.07). This is consistent with regional influences becoming less prominent in favor of global influence under forcing. Further work is required to see whether other regions undergo similar shifts but given the scale of changes in Tmax and RH, similar changes could be expected.

Understanding Pyroclimates
This paper introduces and defines pyroclimates and identifies pyroclimatic regimes. A pyroclimatic regime can be characterized by a steady-state mean daily maximum temperature during the fire season, relative humidity, and regional drought/soil moisture indices. It is distinctly different to fire weather, which is a product of local and short-term conditions constrained by the boundary limits set by the prevailing climate. From this perspective, fire climate shapes fire weather. In a changing climate, fire climate will be an external influence on fire weather, which will in turn be modified by altered local conditions through feedback effects.
The recognition of rapid climate shifts affects the perception of how risks can change over time [40]. Instead of large variations imposed on a background of gradual change, fire risks may change very rapidly. Planning for such changes needs to become a higher priority, especially in areas where a high fire risk jumps from generally low or occasional to being much more frequent. Such areas are less likely to have the resources or experience Anomaly Global fire season length Aust median daily FFDI Figure 9. Global fire season length (the number of days when fire danger was above its median value) [67] compared with the Australian median daily average FFDI [16] for 1979-2013 calendar years.

Understanding Pyroclimates
This paper introduces and defines pyroclimates and identifies pyroclimatic regimes. A pyroclimatic regime can be characterized by a steady-state mean daily maximum temperature during the fire season, relative humidity, and regional drought/soil moisture indices. It is distinctly different to fire weather, which is a product of local and short-term conditions constrained by the boundary limits set by the prevailing climate. From this perspective, fire climate shapes fire weather. In a changing climate, fire climate will be an external influence on fire weather, which will in turn be modified by altered local conditions through feedback effects.
The recognition of rapid climate shifts affects the perception of how risks can change over time [40]. Instead of large variations imposed on a background of gradual change, fire risks may change very rapidly. Planning for such changes needs to become a higher priority, especially in areas where a high fire risk jumps from generally low or occasional to being much more frequent. Such areas are less likely to have the resources or experience to manage such changes. In areas of already high fire risk, fire seasons lengthen, and risks become more chronic, threatening to overwhelm environments and/or change how whole systems function.
As described in Section 3.2.4, the detection and attribution of regime shifts differs from situations where change is trend-like. Baselines for trend detection are usually nominal, with a beginning and end selected on the basis of data availability. They are used for reference periods with the knowledge that more recent baselines are likely to be nonstationary. A climate regime constitutes a physical baseline that remains stationary until a critical threshold is reached and a shift occurs. Using previous climate data to assess fire danger will underestimate risk. Accordingly, in Victoria, the government department managing fire risk on public lands has relied principally on post-1997 fire conditions for some time (Liam Fogarty, pers. comm.). However, future changes are difficult to diagnose due to the inability of climate models to capture changes in relative humidity [47].
Trend analysis can underestimate change for decades following a shift. In Figure 10, modelled Days VHi+ from 1957-2021 for Victoria are shown with trends calculated for five-year intervals from 1957-1991. Trends through to 2001 remain constant and do not respond until around 2006, but only capture half of the actual shift. They retain a constant trend afterwards, not attaining parity with the new regime until 20 years after the shift occurred. In reality, FFDI data did not become available until 2006 and extended back to 1972 [17], so most trends have been calculated from that time. However, a 1972 start using this example would not reach parity with Regime 2 for another decade. In most cases, measuring trends in the presence of regime shifts will underestimate changing risks. respond until around 2006, but only capture half of the actual shift. They retain a constant trend afterwards, not attaining parity with the new regime until 20 years after the shift occurred. In reality, FFDI data did not become available until 2006 and extended back to 1972 [17], so most trends have been calculated from that time. However, a 1972 start using this example would not reach parity with Regime 2 for another decade. In most cases, measuring trends in the presence of regime shifts will underestimate changing risks. This problem compounds with projections of future change, especially if those changes are considered as the continuation of a trend. For example, the change in Figure  10 is at the higher end of projections for 2030 (2015-2045 average) from CCIA2015 [35]. In the previous report, we analyzed changes for all regions, finding that some regions narrowly exceeded the 2030 projections, most were in the upper part, and only South Australia was at the lower end. The Climate Projections Roadmap for Australia that outlines the process of Australia's next set of projections has just been released [68]. It does not outline how evolving climate and risks will be characterized, but more broadly aligns itself with the findings of the Intergovernmental Panel on Climate Change's Sixth Assessment Report [9], which is focused on the construction of model consensus around trendlike change with variability superimposed. This problem compounds with projections of future change, especially if those changes are considered as the continuation of a trend. For example, the change in Figure 10 is at the higher end of projections for 2030 (2015-2045 average) from CCIA2015 [35]. In the previous report, we analyzed changes for all regions, finding that some regions narrowly exceeded the 2030 projections, most were in the upper part, and only South Australia was at the lower end. The Climate Projections Roadmap for Australia that outlines the process of Australia's next set of projections has just been released [68]. It does not outline how evolving climate and risks will be characterized, but more broadly aligns itself with the findings of the Intergovernmental Panel on Climate Change's Sixth Assessment Report [9], which is focused on the construction of model consensus around trend-like change with variability superimposed.
This disjunct between trends, temporal baselines, and projections is partly managed by moving the baseline forward with each iteration, which will take up some of the past change. However, selecting a physically defined baseline and calculating projections as the likelihood of future regime shifts would be more desirable, but currently there is little appetite for changing the status quo.
For fire, overcoming these barriers requires a better understanding of the conditions that define a pyroclimatic regime. Pyroclimates, seasonally wet and dry, are situated towards the drier end of the hydrological cycle. They co-exist with rainfall-runoff and flood cycles, making up the whole wet and dry spectrum. Catchments exhibit runoff regimes that reflect a given rainfall-evaporation regime, where floods are analogous to the wildfire-fire climate relationship. Simple changes in precipitation and potential evaporation can be used to estimate how runoff changes under a changing climate [5], but the nonlinearity of floods makes scaling much more challenging. Both fire and flood events tend to follow a power law in terms of scale and severity. As part of the same overall wet and dry components of the hydrological cycle, it may also be possible to estimate how pyroclimates change by scaling key variables such as fire season T max and 3 p.m. RH. This would require aligning spatial, high-quality meteorological data and point and areal estimates of FFDIs. However, the inability of climate models to reproduce historical shifts in relative humidity and their relative insensitivity to increased forcing [47] suggests some fundamental improvements are required.
Step changes in flow regimes in fire-prone regions in the southern part of Australia are centered on the mid to late 1990s at around the time of the first regime changes in Victoria. These changes were in catchments surrounding the coast and extending a few hundred km inland [69]. Further research may be able to relate both the wet and dry aspects of the hydrological cycle to large-scale regimes shifts.
Paired bivariate tests, especially for the eastern states, showed P, T max , and T max FS increasing relative to each other during the period 2008-2013 without influencing the existing FFDI regimes, except for Queensland which shifted up. This coincided with floods in eastern Australia in 2010-2011 and 2020-2022, interspersed by high fire-danger years. It points to warmer and wetter conditions following the Millennium drought, which have contributed to very high variability in FFDIs. This may be due to warmer sea surface temperatures leading to a more active hydrological cycle and higher land temperatures combining with continuing lowered relative humidity drying out the land surface faster. The potential for large-scale regime shifts to accelerate risk at both extremes of the hydrological cycle needs to be investigated further through process studies analyzing both observations and models.

Conclusions
This paper introduces the concept of a fire climate (pyroclimate), defined as the incoming climate external to a region that governs the mean and distribution of fire weather, thus affecting the propensity for wildfire to occur. A fire climate regime differs from a fire regime that accounts for biomass production, biomass readiness to burn, fire weather, and ignition sources typical of a place or region [3,4]. We have also shown that fire climates form steady-state regimes, similar to those observed for variables such as temperature, rainfall, and relative humidity. Shifts in fire climates occur in response to regime changes in atmospheric heat and moisture content. Identifying and predicting such shifts is therefore central to gaining a better understanding the relationship between climate and fire risk.
We constructed fire climates from the Forest Fire Danger Index which, in Australia, is used to measure fire danger in forested areas and to represent fire danger more widely. The initial model was constructed from a publicly available input of high-quality data from the BoM regressed against annual measures of the FFDI from climate stations across Victoria from 1972-2010. Variables used in the regression were fire season T max , fire year annual rainfall anomaly, the percentage of the region above the 90th percentile of T max , and calendar year 3 p.m. cloud, calculated from T max and P after 2015. Variables produced were ΣFFDI, Days Hi+, Days VHi+, and Days Sev+.
The outputs were compared with an updated record of the FFDI from 1971-2017, specifically median daily FFDIs and the 97th percentile, averaged into fire years (LH2019). The Nash-Sutcliffe efficiency for the model in reproducing the training data for ΣFFDI 1972-2009 was 0.97, and between ΣFFDI and MFFDI 1971-2017, two slightly different measures, was 0.88. The fire climate model was then applied to eight other regions around Australia and compared with station-averaged data from LH2019 for those regions. The Nash-Sutcliffe efficiencies (shown in Table 5) are similar for regions with an even spread of stations (0.84 to 0.88), less so with biased distributions, and are very poor for SW Western Australia. This is a problem with the station data, where the three stations do not agree with each other nor with the regional average climate. High levels of agreement between regions with good station coverage (Victoria, New South Wales, SE Australia, Queensland, and the Northern Territory) show that regions other than Victoria, which provided the training data, can produce results with equivalent skill. The applicability of the model across diverse climates suggests it has a physical basis.
Commencing in in 1995-1996, fire climate regimes began to shift in fire-prone areas of Australia, starting in Victoria and Tasmania and expanding west to South Australia, SW Western Australia and Western Australia, and north to New South Wales through to 2002-2003, affecting Queensland a decade later and potentially the Northern Territory over the period 2011-2017. Fire risk as measured by four indices of the FFDI has shifted by: • aproximately one-quarter in the SE to 8% in the west for ΣFFDI; • approximately one-third in the SE to 7% in the west for Days Hi+; • approximately half in the SE to 11% in the west for Days VHi+, with a greater increase in Tasmania; • approximately three-quarters in the SE to 9% in the west for Days Sev+, with no result for Tasmania.
Detailed changes showing dates and changes in mean and p values are shown in Figure 11. These estimates are likely to be conservative, given the model underestimates regime shifts for Victoria where the model was constructed, by approximately 15% relative to the size of the shift in the baseline data. The attribution of change using trend analysis can substantially underestimate the impact of regime changes ( Figure 10). For externally forced regime changes that shift from one steady-state to another, the total change component can be attributed to that forcing. This improves our ability to predict the return periods of both fire seasons and extreme fire weather significantly. It is far more accurate than using trend analysis but remains relatively imprecise. Ranges of confidence will remain high because of the relatively short periods available for analysis but can be narrowed with further work.
Estimates of the Regime 2 mean as a Regime 1 percentile range from the 0.79 percentile for Tasmania and 0.80 for the Northern Territory and up to 0.94 to 0.97 of Regime 1 for Queensland, averaging 0.90. A shift from the 50th to 90th percentile is substantial. We also assessed the factors of exceedance by Regime 2 the 1:10 and 1:20 fire seasons from Regime 1. The 1 in 10 fire season of Regime 1 is exceeded three to seven times as often in Regime 2 with an average factor of five. The 1 in 20 fire season is exceeded from twice to 14 times more often, averaging 8 times more often. The upper limit of Regime 1 is exceeded in 25% and 50% of fire years in Regime 2 depending on location. These are sobering results. Similar changes may well be occurring in other fire-prone regions of the world but have not been fully recognized.
Sensitivity testing and paired bivariate analysis identified the drivers of the regime shifts as fire season maximum temperature and 3 p.m. relative humidity, with their timing and magnitude shaped by drought factor/soil moisture and rainfall. Regime shifts in Tmax and RH were globally widespread in the period 1997 to 2003 [45,47]. Consistent with these changes, median daily FFDIs for Australia and global fire season length calculated from The attribution of change using trend analysis can substantially underestimate the impact of regime changes ( Figure 10). For externally forced regime changes that shift from one steady-state to another, the total change component can be attributed to that forcing. This improves our ability to predict the return periods of both fire seasons and extreme fire weather significantly. It is far more accurate than using trend analysis but remains relatively imprecise. Ranges of confidence will remain high because of the relatively short periods available for analysis but can be narrowed with further work.
Estimates of the Regime 2 mean as a Regime 1 percentile range from the 0.79 percentile for Tasmania and 0.80 for the Northern Territory and up to 0.94 to 0.97 of Regime 1 for Queensland, averaging 0.90. A shift from the 50th to 90th percentile is substantial. We also assessed the factors of exceedance by Regime 2 the 1:10 and 1:20 fire seasons from Regime 1. The 1 in 10 fire season of Regime 1 is exceeded three to seven times as often in Regime 2 with an average factor of five. The 1 in 20 fire season is exceeded from twice to 14 times more often, averaging 8 times more often. The upper limit of Regime 1 is exceeded in 25% and 50% of fire years in Regime 2 depending on location. These are sobering results. Similar changes may well be occurring in other fire-prone regions of the world but have not been fully recognized. Sensitivity testing and paired bivariate analysis identified the drivers of the regime shifts as fire season maximum temperature and 3 p.m. relative humidity, with their timing and magnitude shaped by drought factor/soil moisture and rainfall. Regime shifts in T max and RH were globally widespread in the period 1997 to 2003 [45,47]. Consistent with these changes, median daily FFDIs for Australia and global fire season length calculated from model reanalysis both shifted in the calendar year for 2002 ( Figure 9). This indicates that fire danger globally is responding to large-scale forcing involving regime shifts in sensible and latent heat. Fire climates can therefore represent the external forced component of changing fire danger.
Fire risk in a warming climate is complex due to changing fuel amounts, types, moisture, ignition sources, and fire weather. This paper introduces the concept of fire climates to describe the boundary conditions that govern the overall distribution of fire weather. It provides a way to estimate the externally forced component of changing fire risk. Further work is needed to develop a working model for understanding how fire risk is likely to respond to ongoing climate change, not least by applying the concept to other parts of the world and exploring how fire climates are evolving in the current generation of earth system models. Acknowledgments: Thanks to Chris Lucas of the Australian BoM who supplied the original data FFDI in 2010. Circumstances meant the initial paper describing the adjustments and regime changes was never completed. Thank you also to the Victorian Department of Energy, Environment and Climate Action, formerly Environment, Land, Water and Planning, who have accepted the reality of regime shifts in fire climates. The pers. comm. in the paper is from Liam Fogarty who was involved in strategic planning as a Senior Policy Officer. Thanks also to the fellow members of Victoria's bushfire risk modeling system and risk-based decision-making panel and to those in public service who engaged with the previous report, Constructing and Assessing Fire Climates for Australia. These discussions helped focus the current paper. Data from Lucas and Harris [16] and Jolly et al. [67] are both provided under the CC BY 4.0 Licence.

Conflicts of Interest:
The authors declare no conflict of interest. Table A1 shows the constants for the HQD regression model. The p values produced are also shown. The intercept shows high p values, suggesting it is not essential for prediction, but it is needed to set a minimum value. Some of the variables increase above p = 0.05 but the sensitivity testing in Tables A2 and A3 show they are essential to the overall model. The full results are in the Supplementary Materials.

Appendix A. Regression Results and Sensitivity Testing
For Table A2, each variable was omitted from the regression and the results tallied. The main outputs show the variation to the mean, shift size and timing, and standard deviation. They show different behavior within and between FFDIs as described in Section 3.1.3. FFDI results have also been allowed to fall below zero in response to the removal of T max FS as a sensitivity exercise. Table A3 shows the results for changes that are closer to what might be experienced by observed changes. The variables were each varied by 10%, whereas the standard deviations for P from 1957-2021 are 15%, for T max FS 3%, T max 90 is almost 200% due to the highly nonlinear time series, and C3pm is 5%.

Appendix B. Nonlinear Attribution Test Results
This Appendix contains results for the paired bivariate tests. Due to their novelty, we have included a comprehensive set of results for each region tested. For each table, the upper part of each panel showing single variables was tested against a random reference time series, which is the standard way of assessing regime shifts. The lower part contains paired shifts where one is the reference and the other is the test, and these are mirrored.
The following points will help interpret those results:

•
Two time series containing a common shift will show no shift when paired if those shifts are proportional. For example, pairing the different measures of FFDIs for a single region shows no regime change when paired or mirrored (p > 0.10). This may denote a common origin for both variables or direct cause-and-effect where one forces the other (i.e., full attribution). • Partial attribution may be the case when a pair has similar regime shifts and one shows no shift when paired in one direction, but when mirrored only part of the original shift can be attributed. RH shows this for many regions: the FFDI shows no shift against RH, but RH against the FFDI reproduces only part of its regime change. For example, for Victoria, RH shifted by −3.2 in 1996-1997 which is compatible with the FFDI, but in reverse it showed a relative shift of −1.4 in the same year, so only −1.8 can be accounted for. • Shifts due to an independent process may show conformity in one direction but none when mirrored. This is consistent with windspeed, which shows a potential attribution of 0.1−0.3 m s −1 for regime shifts in the early to mid-1990s that are in excess of 2.0 m s −1 , where we know measurement methods changed over at that time.

•
When two variables with shifts in common show no shift at that time, but relative shifts at other times, they are conformal for the regime shift but reflect the presence of other processes that will need a separate interpretation.

•
Other results can show nonlinear responses or shifts due to an independent process not directly related to a change in fire climates, including inhomogeneities. Climate variability may also be a factor in short sequences.