Introduction

Wildfire can have severe environmental and societal impacts, affecting ecosystems, agricultural lands and urban settlements1,2,3,4. Recent high-profile disasters, such as those in Chile and Portugal in 20175,6, Greece in 20187, California in 2018 and 20208,9, and Australia in 2019–2010 highlight the ongoing need for wildfire adaptation and mitigation strategies11.

Variations in weather and climate influence wildfire activity by modulating vegetation production and fuel aridity12. Precipitation deficits induce moisture stress, increasing the flammability of fuel13,14,15. Yet the relationship between drought and wildfire varies markedly, with distinct differences dependent on the ecosystem15,16. In grasslands and savannas, drought is associated with decreased wildfire potential3,17,18. Antecedent precipitation promotes vegetation growth, yielding an abundance of available fuel that becomes flammable in the dry season. In these ecosystems, then, drought during the wet season results in a fuel-limited environment, suppressing the ability of wildfire to spread15. It is these frequent, low-intensity wildfires that account for the majority of burned area globally19.

Conversely, in forest biomes, drought is associated with increased wildfire activity through a reduction in fuel moisture content3,14,17,18,20,21,22,23. Pertinently, the majority of economically or socially disastrous wildfires occur in temperate and boreal forest3. This suggests that drought plays a role in the costliest wildfire events. Multi-year drought was a major contributing factor to the severity of the 2019–20 wildfire season in Australia24,25,26,27. Wet-season precipitation deficits, combined with above-average fuel loads, have been implicated in the 2018 California wildfires8.

Fire weather, which is characterised by high temperatures, low humidity and strong winds, is an important driver of wildfire across much of the globe3,28,29,30,31,32,33,34. As a product of multiple processes, fire weather is a form of compound event. Wind plays a crucial role in determining wildfire behaviour, affecting fire spread, intensity and ignition likelihood15. Wind, temperature and relative humidity can exacerbate flammability by warming and drying fuels15. Together with recent precipitation, these variables are often combined as fire weather indices, which are used routinely in wildfire management35,36. In many wildfire disaster-prone regions, such as western North America, the Mediterranean and southeast Australia, fire weather has become more frequent and more severe34,37,38,39,40,41,42,43,44,45.

Moreover, anthropogenic climate change has extended the length of the fire weather season for many regions39,46,47,48,49. Future projections indicate this trend will continue40,42,50,51,52,53. This poses challenges for managing firefighting resources across regions that historically have distinctly separate wildfire seasons. The USA, Canada, Australia and New Zealand cooperate through shared firefighting resources such as equipment and personnel10,54,55. In the 2019–20 Australian fire season, hundreds of overseas firefighters were deployed to assist operations, and 11 of the 12 large air tankers used were sourced internationally10. In response, the Australian Royal Commission into National Natural Disaster Arrangements10 specifically highlighted that lengthening fire seasons in the northern and southern hemispheres may reduce the availability of aircraft services to Australia. The severity of the following USA wildfire season further underlines these concerns. In 2020 major wildfires continued well through boreal autumn and into winter56,57. The season ended with over four million hectares burned, including more than 4% of California9,58,59. Fortunately, the 2019–20 wildfires in Australia and the subsequent 2020 USA wildfires were separated by a number of months. Concurrent wildfires of similar severity to those events could place significant pressure on international cooperation.

The interaction of drought and fire weather constitutes a “compound event”. The field of compound events has gained significant research interest in recent years, with the recognition that a combination of multiple hazards or drivers can amplify impacts60,61,62,63,64,65,66,67.

There are numerous studies on “hot and dry” compound extremes, the majority of which focus on precipitation deficits and high temperatures over a common period, for example hot and dry summers43,68,69,70,71,72,73,74,75,76,77. Consideration of hazards over different time scales is less common, although some studies assess the joint behaviour of daily temperature extremes and seasonal precipitation66,78,79. Furthermore, the studies that explicitly consider fire weather and drought only consider the latter on time-scales of at most 3 months43,66,74. This neglects the long-term processes that can be associated with increased wildfire potential in some regions. An exception to this is a study that assessed the role of annual precipitation and monthly fire weather during the Australian 2019–20 fire season34.

Although the role of drought and extreme fire weather in driving wildfire potential is well known, the preceding discussion highlights that these hazards are often considered independently. Furthermore, many such analyses are regional, and do not consider the possibilities of synchronous wildfire potential in multiple regions. The extent to which extreme fire weather and drought compound, and how their co-occurrence has changed through time, has not been explored on a global scale.

In this article, we explore the relationship between global fire weather and long-term antecedent meteorological drought. Using the Forest Fire Danger Index (FFDI, see “Methods”) and accumulated precipitation, we quantify the co-occurrence of these two hazards, identify regional “hot spots” and trends, and analyse potential climate drivers. We follow the definitions provided in a recent typology of compound events to differentiate between sub-classes of compound events67. A hazard that causes or amplifies an impact only as a consequence of a preceding hazard is known as a “preconditioned event”. Here, meteorological drought is considered the precondition that might amplify the impact caused by a period of heightened fire weather.

We also discuss the simultaneous occurrence of these hazards across different regions. Hazards that occur in different regions within a time frame so as to cause an impact are classed as “spatially compounding events”67. We address the growing concerns of lengthening fire weather seasons in the USA and Australia by analysing changes in spatially compounding fire weather and drought in these nations. We will also explore possible drivers of these events by considering a range of climate modes associated with fire weather in these regions.

Results

Relationship of burned area to fire weather and antecedent precipitation

The FFDI is highest on average in desert regions that do not feature wildfires (Fig. 1a, b). This is not surprising given that temperature and relative humidity are components of the index. There is a clear correspondence between the seasonality of burned area and FFDI (Fig. 1c, d). The months that, on average, yield the maximum burned area and FFDI are often the same. Otherwise, in places such as southern Australia and western North America, the maximum FFDI leads the maximum burned area by a month or more, suggesting that wildfires are most widespread following the peak of the fire weather season.

Fig. 1: FFDI and burned area climatology 2001–2019, and event definitions.
figure 1

a Average annual burned area (km2). b Average daily FFDI. c Month of maximum average burned area. d Month of maximum average FFDI. e Example of which months are used in the calculation of annual values of precipitation, burned area and fire weather days (FWD). The data are for the eastern Australia (EAU) region shown in (a). October is the month of average maximum FFDI. Therefore the 2020 event is defined using the period November 2018 through October 2019 for precipitation, and April 2019 through March 2020 for FWD and burned area.

Full size image

We define a “fire weather day” (FWD) as when the FFDI exceeds the climatological 95th percentile. This threshold is calculated using all data (i.e., not as a function of month or season), so the vast majority of fire weather days occur during the fire season45. The number of fire weather days accumulated over some period, L, is denoted FWDL. We similarly estimate drought using accumulated precipitation, PL.

We consider annual statistics to analyse preconditioned drought and fire weather events. Using the peak FFDI months shown in Fig. 1d, we target the 12-month period best suited to estimating the severity of the fire weather season and the magnitude of a preceding drought. For annual fire weather, FWDL, we set L = f, where f denotes the “fire year”. This is the 12-month period from six months prior to five months following the peak FFDI month. We also utilise satellite-derived burned area data, which we accumulate over the same period and denote as BAf. For annual precipitation, PL, we choose an antecedent period, L = a. This period encompasses the 12 months leading up to and including the peak FFDI month, providing an estimate of fuel dryness at the peak of the fire weather season. Therefore, the extent to which a “year” can be classified as a preconditioned event includes data from an 17-month period, and is labelled according to the year of the last month in that period.

An illustration of these quantities is shown in Fig. 1e, for the eastern Australia (EAU) region of the Intergovernmental Panel on Climate Change (IPCC) regions80. For denoting spatially-averaged quantities, we use a superscript indicating the region, r. The month of average maximum FFDIEAU is October. Therefore, antecedent precipitation, \({P}_{a}^{{{{\rm{EAU}}}}}\), is accumulated for November through October. Fire weather days, \({{{{\rm{FWD}}}}}_{f}^{{{{\rm{EAU}}}}}\), are accumulated for April through March. The example in Fig. 1e, for the 2020 event, highlights the lack of precipitation leading up to the fire season, particularly during the preceding winter. Also shown is the high precipitation in February 2020 that was critical in extinguishing many of the wildfires, with the burned area reducing to virtually zero. The number of fire weather days was high throughout the warm season. This peaked in December, with all but one day classed as a fire weather day (also found for a region southward of EAU by another study45).

We find that annual variations in fire weather exert greater influence on burned area than variations in antecedent precipitation. Years with anomalously high numbers of fire weather days correspond to greater burned area anomalies than years of anomalous antecedent precipitation (Fig. 2a–c). There is a positive relationship between burned area and fire weather for much of the globe, including North and South America, Europe and large swathes of Asia (Fig. 2b). The relationship between accumulated precipitation and burned area is more varied. Dry conditions appear to be a factor in increased burned area for tropical forested regions such as the Amazon and the Maritime Continent (Fig. 2a), in agreement with other studies21,81,82. Increased precipitation prior to the peak fire weather season is associated with positive burned area anomalies in much of Australia and parts of Africa and the Mediterranean (Fig. 2c), due to the promotion of fine fuel growth3,17,18. Regarding Australia, note that in the southeast coastal strip, where the majority of national wildfire disasters occur, burned area anomalies are positive for dry antecedent conditions and for higher numbers of fire weather days. The short burned area record (20 years) means these results should be treated cautiously. In particular, a permutation test shows that the results described here for dry and wet years are not statistically significant for many grid cells (Supplementary Figure 1).

Fig. 2: Relationship of burned area to fire weather and precipitation.
figure 2

a Annual normalised burned area anomalies, \(({{{{\rm{BA}}}}}_{f})^{\prime} /{\sigma }_{{{{{\rm{BA}}}}}_{f}}\) (dashes indicate anomalies), for 2001–2019, averaged over dry antecedent years, defined as when \({P}_{a} \,<\, {\mu }_{{P}_{a}}-{\sigma }_{{P}_{a}}/2\). For most grid cells, five to seven years satisfy this threshold. b As in a but for years of heightened fire weather, defined as when \({{{{\rm{FWD}}}}}_{f} \,>\, {\mu }_{{{{{\rm{FWD}}}}}_{f}}+{\sigma }_{{{{{\rm{FWD}}}}}_{f}}/2\). c As in a but for wet antecedent years, \({P}_{a} \,>\, {\mu }_{{P}_{a}}+{\sigma }_{{P}_{a}}/2\). d Grid cells for which burned area anomalies increase when wet (blue; c) or dry (red; a) years compound with years of heightened fire weather (b). White indicates no increase in burned area from compound dry or wet years. There are typically only one to three of these compound years, but results are robust when computed with smaller departures from the mean e.g. using μ ± σ/4.

Full size image

For most of the globe, burned area during years with anomalously high fire weather is increased when the preceding year has been drier than average (Fig. 2d). In a few regions, such as northwest Australia and Madagascar, heightened fire weather preceded by a wetter-than-average year is associated with increased burned area. This highlights the importance of considering compound climate extremes in relation to wildfire potential. As wildfire potential in most regions is exacerbated by heightened fire weather compounding with antecedent dry conditions, our focus for the remainder of the paper is on compound drought and fire weather.

Compound drought and fire weather events

We define a “drought year” as when accumulated precipitation is less than one standard deviation below the mean, \({P}_{a} \,<\, {\mu }_{{P}_{a}}-{\sigma }_{{P}_{a}}\). Similarly, a “fire weather year” is defined as when \({{{{\rm{FWD}}}}}_{f} \,>\, {\mu }_{{{{{\rm{FWD}}}}}_{f}}+{\sigma }_{{{{{\rm{FWD}}}}}_{f}}\). There are clear “hot spots” in which preconditioned drought and fire weather years occur (Fig. 3a). This includes regions that suffer from wildfire disasters, such as western North America, the Mediterranean and southeast Australia. Furthermore, in these regions the number of preconditioned events has increased since 1996 (Fig. 3b).

Fig. 3: Climatology and changes in preconditioned drought and fire weather.
figure 3

a Number of preconditioned drought \({P}_{a} \,<\, {\mu }_{{P}_{a}}-{\sigma }_{{P}_{a}}\) and fire weather \({{{{\rm{FWD}}}}}_{f} \,>\, {\mu }_{{{{{\rm{FWD}}}}}_{f}}+{\sigma }_{{{{{\rm{FWD}}}}}_{f}}\) years 1970–2020. b Change in number of preconditioned drought and fire weather years from 1971-1995 to 1996-2020. c–h Annual number of fire weather days, \({{{{\rm{FWD}}}}}_{f}^{r}\) (orange), and antecedent precipitation, \({P}_{a}^{r}\) (blue), for six IPCC regions. The data are scaled to have unit variance and adjusted such that the threshold for drought (\({\mu }_{{P}_{a}^{r}}-{\sigma }_{{P}_{a}^{r}}\)) and fire weather (\({\mu }_{{{{{\rm{FWD}}}}}_{f}^{r}}+{\sigma }_{{{{{\rm{FWD}}}}}_{f}^{r}}\)) years are the same. The y-axis for precipitation is inverted such that both drought and fire weather years lie above the threshold (the black horizontal line labelled μ ± σ). Preconditioned drought and fire weather years are indicated by vertical grey lines. i-k Proportion of area in drought, fire weather, and preconditioned years, respectively. Colours indicate data for the Northern Hemisphere, the Southern Hemisphere, the tropics, and global.

Full size image

Using six IPCC regions as an illustration, we find that increases in the frequency of preconditioned events appear to be driven primarily by changes in fire weather rather than drought (Fig. 3c–h). In western North America (WNA), there have been eight fire weather years since 1990, but only two of these coincide with drought. Similarly, a steep rise in fire weather days in western South Africa (WSAF) since 2000 is mitigated by increased precipitation, hence only two preconditioned event years. The Mediterranean (MED), on the other hand, has seen six preconditioned event years since 2000, driven by declining precipitation and an apparent step change in the frequency of fire weather days. A possible step change is also evident in southern Australia (SAU). Siberia has experienced significant wildfires in recent years83, and we find record high numbers of fire weather days in eastern Siberia (ESB) in 2017 and 2018. Increased wildfire potential from an upward trend in fire weather days is, however, moderated by increased precipitation.

To further investigate the nature of these changes, we apply statistical tests for monotonic trends and change points in precipitation and fire weather (“Methods”). Although annual precipitation has changed for some regions, we do not find sufficient statistical support for trends (Fig. 4a) or step changes (Supplementary Fig. 3) in most regions. The exception to this is in parts of high-latitude Europe and Asia, which show significant increases in precipitation of up to 0.5 standard deviations per year. For fire weather days, we find evidence of increasing trends and change points for many parts of the world, including western North America, Brazil, eastern Europe and western Asia, and southeast Australia (Fig. 4b–d). Furthermore, step changes in several key regions (western North America, the Mediterranean and southeast Australia) occurred at around the same time, in the 1990s to early 2000s (Fig. 4c), in agreement with findings from regional studies37,45.

Fig. 4: Fire weather and precipitation trends and change points 1970–2020.
figure 4

a Theil-Sen slope for normalised annual antecedent precipitation anomalies, \(P^{\prime} /{\sigma }_{{P}_{a}}\) (dashes indicate anomalies). b As in a but for annual numbers of fire weather days, FWDf. Hatching in (a) and (b) indicates a statistically significant result for the Mann-Kendall test for monotonic trends. c Pettitt test year of potential step change in annual numbers of fire weather days FWDf. d Pettitt test step change magnitude for FWDf. Hatching in (d) indicates statistical significance for the Pettitt test.

Full size image

For regions that now experience more frequent occurrences of fire weather days, the change appears to be driven by higher daily maximum temperatures and lower daily relative humidity (Supplementary Fig. 4). Decreases in fire weather days, for example in India and northeastern Russia, appear to be driven primarily by increased relative humidity, which overrides increases in maximum daily temperature.

The dominant role of the FFDI in driving widespread increases in preconditioned drought and fire weather events is also evident globally at the grid-cell scale. The proportion of area in drought in any given year has remained steady since 1959, perhaps undergoing a slight decrease (Fig. 3i). The number of fire weather days, on the other hand, has increased since 1980 (Fig. 3j). These differing trends combine to result in an increase in preconditioned events (Fig. 3k). In 2019–20, drought afflicted almost 50% of the Southern Hemisphere, with almost 70% affected by extreme fire weather conditions. A remarkable 42% suffered both drought and fire weather years in 2019–20, driven by how widespread the severe conditions were in Australia.

Drivers of preconditioned and spatially compounding events

We composite a number of atmospheric and oceanic diagnostics on the 10% of years (six years) that featured the most widespread preconditioned drought and fire weather conditions (Fig. 3i–k). Due to the contrasting fire weather seasons, we consider the Northern Hemisphere, the tropics and the Southern Hemisphere separately.

In the Northern Hemisphere, the six most widespread preconditioned years feature precipitation deficits and more numerous fire weather days across western-central Russia and Kazakhstan, eastern Asia and southern and eastern North America (Fig. 5a, b). Pacific SST anomalies during the 12 months prior to the peak of the fire weather season resemble La Ni\(\tilde{{{{\rm{n}}}}}\)a and negative Pacific Decadal Oscillation (PDO) signatures (Fig. 5a), consistent with the regional relationships between these modes and precipitation84,85,86,87,88,89. Large-scale anomalies of velocity potential at 200 hPa (ϕ200; Fig. 5a) indicate associated changes in the Walker circulation, which in turn is linked to reduced precipitation in southern North America and increased precipitation over Australia. During May, June and July of these years, positive 500 hPa geopotential height (z500) anomalies suggest atmospheric ridging or blocking played a role in the heightened fire weather in northwestern Asia and North America (Fig. 5b). This is consistent with other studies: ridging is a feature of North America fire spread events90 and increased wildfire potential91,92,93,94, while a long-lived block was implicated in the 2010 western Russia heatwave and wildfires95,96.

Fig. 5: Composites of spatially extensive preconditioned drought and fire weather years.
figure 5

a Normalised antecedent precipitation (\({P}_{a}/{\sigma }_{{P}_{a}}\); brown and green shading), sea surface temperature (SST; red and blue shading), 500 hPa geopotential heights (z500; black contours starting at ± 5 m and spaced in intervals of 10 m) and 200 hPa velocity potential (ϕ200; pink contours starting at ± 5e5 m2s−1 and spaced in intervals of 5e5 m2s−1). The data are anomalies averaged over the six years listed to the left of the panel, selected as the 10% of years with the most widespread drought and fire weather in the Northern Hemisphere. b As in (a) but for fire weather day, FWDf, anomalies (shading) and three-monthly z500 anomalies (contours starting at ± 5 m and spaced in intervals of 10 m). c–f As in (a) and (b) but for the tropics and the Southern Hemisphere. In all panels, precipitation and fire weather data are for the “fire year” (L = f) and “antecedent” (L = a) periods, respectively, which vary spatially (see the text and Fig. 1d). The period used for other variables in the left (right) column is the 12 (three) months ending (centred) on the most common peak FFDI month for the Northern Hemisphere (June), tropics (August), or Southern Hemisphere (January). Solid (dashed) contours indicate positive (negative) anomalies.

Full size image

The signal from the z500 anomaly composite, however, suffers from substantial smoothing of local features. Due to the large land area of the Northern Hemisphere (compared to the tropics or the Southern Hemisphere), there is greater variability amongst the members of the composite shown in Fig. 5a, b. The locations of simultaneous fire weather “hot spots” across the hemisphere vary between individual years. The common feature of these events regardless of where they occur is that nodes of positive z500 anomalies are co-located with areas of heightened fire weather (Supplementary Fig. 5). Furthermore, on individual years, analysis of 300 hPa zonal and meridional wind provides evidence of circumglobal wave trains with a wavenumber ranging between five and seven. This suggests that spatially compounding fire weather events may be linked to recurrent Rossby wave patterns. Indeed, it has been shown that Rossby waves with wavenumbers five and seven are linked to simultaneous summer heatwaves across the hemisphere97,98,99.

Preconditioned drought and fire weather years in the tropics have become more widespread over time, with five of the six years with the highest proportion of affected area occurring since 1998 (Figs. 3k and 5c, d). In agreement with other studies21,69,89,100,101,102,103, a strong El Ni\(\tilde{{{{\rm{n}}}}}\)o signature is associated with drought affecting western and central Africa, the Maritime Continent and northeastern South America (Fig. 5c and Supplementary Fig. 6). Positive ϕ200 anomalies over the western Pacific and eastern Indian Ocean indicate upper-level convergence, associated with a weakening of the ascending branch of the Walker circulation. This can result in a weakened boreal summer monsoon and is associated with drought across the tropics103. Indeed drought, more so than fire weather, is regarded as a key climate driver of wildfire potential in tropical rainforest81,82,104,105,106,107.

Southern Hemisphere preconditioned events span southern South Africa and southeast Australia, with a mixed response in southern South America (Fig. 5e, f). A much weaker El Ni\(\tilde{{{{\rm{n}}}}}\)o pattern is evident compared to tropical preconditioned events (Fig. 5e). Analysis of the individual years in the composite reveals a combination of El Ni\(\tilde{{{{\rm{n}}}}}\)o- and La Ni\(\tilde{{{{\rm{n}}}}}\)a-like SST patterns (Supplementary Fig. 7). This is perhaps due to the differing responses to ENSO across the Southern Hemisphere. While El Ni\(\tilde{{{{\rm{n}}}}}\)o tends to be related to drought and extreme fire weather in southeast Australia32,34,108,109, ENSO impacts on precipitation and temperature in southern African and southern South America are multi-faceted, dependent on the stage of the ENSO cycle and on the season102,110,111,112.

Since the tropical regions are not providing a unique pattern to force concurrent extratropical extremes, we look to the large-scale atmospheric circulation to see if it might assume this role. Positive z500 anomalies are evident over Antarctica, with negative anomalies to the north, in the 12-month period leading to the peak fire season (February through January) and the 3-month period centred over this peak (December through February). This feature resembles the negative phase of the Southern Annular Mode113 (SAM), which can be associated with reduced precipitation in southern Australia and southeastern South Africa34,114,115. In southern South America, however, enhanced wildfire activity can be associated with the opposite (positive) phase of the SAM116,117. This is consistent with our findings, which show a reduction in the number of fire weather days in this region (Fig. 5f). Coincident with the negative-phase SAM, we also find evidence of a wave train with wavenumber four that may link coincident fire weather anomalies across the hemisphere (Fig. 5f). Wave trains with wavenumbers four or five are evident in the individual years (Supplementary Fig. 7). This supports previous research which established that these wavenumbers are associated with amplified wave resonance and circumglobal, spatially compounding temperature extremes in austral summer97.

Spatially compounding fire weather in western North America and eastern Australia

In this section we focus on changes in overlapping fire weather seasons in western North America (WNA) and eastern Australia (EAU). We switch to monthly data, calculating three-month accumulations of fire weather days and burned area (denoted \({{{{\rm{FWD}}}}}_{3}^{r}\) and \({{{{\rm{BA}}}}}_{3}^{r}\), respectively), and 12-month accumulations of precipitation (\({P}_{12}^{r}\)). These accumulations are calculated on a rolling basis and are right-labelled according to the last month of the accumulation period. Similarly to the previous section, we consider a fire weather or burned area (drought) event as a month which exceeds (is below) one standard deviation from the mean.

The mid-1970s to the late 1990s was a relatively quiescent period for fire weather in both regions. From 2000 the number of fire weather events dramatically increases (Fig. 6), consistent with trend and step change results for these regions found here (Fig. 4) and in other studies31,33,39,42,45,118,119.

Fig. 6: Fire weather, drought and burned area in western North America (WNA) and eastern Australia (EAU), and related climate modes.
figure 6

Top Fire weather events, \({{{{\rm{FWD}}}}}_{3}^{r} \,>\, {\mu }_{{{{{\rm{FWD}}}}}_{3}^{r}}+{\sigma }_{{{{{\rm{FWD}}}}}_{3}^{r}}\) (rectangles); burned area events, \({{{{\rm{BA}}}}}_{3}^{r} \,>\, {\mu }_{{{{{\rm{BA}}}}}_{3}^{r}}+{\sigma }_{{{{{\rm{BA}}}}}_{3}^{r}}\) (open and filled circles); and drought events, \({P}_{12}^{r} \,<\, {\mu }_{{P}_{12}^{r}}-{\sigma }_{{P}_{12}^{r}}\) (vertical lines with cross markers). Years in which there were at least three fire weather events (FWE) in a region are marked with a triangle or square below. WNA results are in blue, EAU results are in orange, and concurrent results are in pink. Bottom The Ni\(\tilde{{{{\rm{n}}}}}\)o3.4 index, the dipole mode index (DMI), the southern annular mode index (SAM), the Pacific-North America pattern index (PNA) and the Gulf of Alaska ridge index (GAR), averaged over the three months indicated by the last three letters.

Full size image

As a consequence, the two regions are now far more likely to experience prolonged periods of increased fire weather in the same year (Fig. 6). Between 1958 and 2000, there were three years during which WNA and EAU both experienced at least three fire weather events (indicated by the pink squares below the top panel of Fig. 6). Such years occurred eight times in the remaining 20 years from 2001, including three consecutive years between 2017 and 2019. Moreover, there has been an increase in the number of overlapping (concurrent) fire weather events across regions, with one occurrence prior to 2000, and seven after 2000 (pink rectangles in the top panel). These overlaps typically occur in October, but the Septembers of 2017 and 2019 are also concurrent events. The prolonged USA fire weather season of 2020 is evident, with five consecutive fire weather events, including the first such November on record. These results complement a previous study, which showed that between 1973 and 2013 the length of fire weather seasons in California increased, and the frequency of long fire weather seasons in eastern Australia and western North America has increased over the same period48.

Spatially compounding fire weather in these two regions is associated with El Ni\(\tilde{{{{\rm{n}}}}}\)o-like SSTs in October through December. This is evidenced by the positive Ni\(\tilde{{{{\rm{n}}}}}\)o3.4 values for eight of the 11 years featuring at least three fire weather events in both regions (pink squares; Fig. 6). In eastern Australia, years of increased wildfire potential are typically associated with El Ni\(\tilde{{{{\rm{n}}}}}\)o32,34,108,120. Accordingly, we find that six of nine years with at least three fire weather events in EAU (orange triangles) feature an El Ni\(\tilde{{{{\rm{n}}}}}\)o-like pattern. On the other hand, of the 14 such years in WNA, nine feature La Ni\(\tilde{{{{\rm{n}}}}}\)a-like SSTs (blue triangles). The mixed response of WNA fire weather to ENSO accords with other regional studies, which show substantial differences across the western USA121,122,123,124,125.

In addition to El Ni\(\tilde{{{{\rm{n}}}}}\)o, we find that years of increased fire weather days in EAU tend to be associated with a positive phase IOD and negative phase SAM (Fig. 6). The role of these three modes in promoting Australian wildfire potential is well established32,34,108,120. In particular, it is the co-occurrence of fire-promoting phases of these modes that may be crucial, rather than their magnitudes34. Our findings accord with this—while the majority of years with at least three EAU fire weather events feature these modes in their fire-promoting phase, their averaged magnitudes are not large (no greater than ∣0.8∣) and are not statistically distinct from zero.

The Pacific North America (PNA) pattern126 is associated with fluctuations in the East Asia jet stream. In its positive phase, the PNA is associated with increased temperatures and reduced precipitation in western North America127,128,129. Similar circulation characteristics have been identified during large fire years93 and years of increased climate-related wildfire potential92. However, we find no evidence of this, with years featuring at least three WNA fire weather events just as likely to feature a negative winter PNA index as a positive index (Fig. 6). Due to the association of local ridges with fire events in the region43,90, we use the Gulf of Alaska ridge (GAR) index130,131 to indicate anomalous upper-level high-pressure activity. We find that boreal autumn (ASO) values of the GAR index are generally positive during years with at least three WNA fire weather events, suggesting enhanced anticyclonic circulation and meridional shifts in the jet stream.

Discussion

We have analysed changes in wildfire potential related to compound fire weather and meteorological drought. Regions prone to wildfire disasters, including southern Australia, the Mediterranean, and the western USA, are among the most likely to experience fire weather years preconditioned by drought. Furthermore, much of the globe has experienced an increase in such events, driven by changes in fire weather rather than precipitation.

Our analysis shows that the likelihood of spatially compounding fire weather in western North America (WNA) and eastern Australia (EAU) has increased. In both regions, the number of fire weather events (three-month periods with anomalously high numbers of fire weather days) has increased substantially in the past two decades. As a result, these regions have become more likely to experience prolonged periods of heightened fire weather, with a greater likelihood of the fire weather seasons overlapping.

ENSO is a key driver of global drought and fire weather. Its warm phase, El Ni\(\tilde{{{{\rm{n}}}}}\)o, is strongly associated with preconditioned events in the tropics, most likely due to its moderating effects on monsoonal precipitation103. An El Ni\(\tilde{{{{\rm{n}}}}}\)o signature is comparatively weaker for Southern Hemisphere preconditioned events, emerging as a composite from both El Ni\(\tilde{{{{\rm{n}}}}}\)o- and La Ni\(\tilde{{{{\rm{n}}}}}\)a-like SST patterns, in line with the differing continental responses to ENSO. The key feature associated with drought and fire weather in the Southern Hemisphere is a negative SAM-like pattern. This pattern is evident throughout both austral summer and the preceding year, and is consistent with reduced precipitation and heightened fire weather in southern Australia and South Africa. In contrast, Northern Hemisphere preconditioned events are associated with La Ni\(\tilde{{{{\rm{n}}}}}\)a. Atmospheric ridging and blocking is also evident at higher latitudes in boreal summer, supporting previous research highlighting the importance of jet stream variability in modulating wildfire activity90.

Although La Ni\(\tilde{{{{\rm{n}}}}}\)a tends to coincide with increased fire weather in the Northern Hemisphere, spatially compounding events in WNA and EAU are associated with El Ni\(\tilde{{{{\rm{n}}}}}\)o. The average magnitudes of ENSO and other, more local, climate modes are small for these events. It is possible that the co-occurrence of fire-promoting phases of these modes is important, rather than their individual magnitudes32,34. Alternatively, the role of the modes discussed here may be relatively inconsequential in relation to spatially compounding fire weather events.

Our finding that upward trends in climate-related wildfire potential is driven by increasingly severe fire weather likely reflects the impacts of anthropogenic climate change. Higher temperatures are sufficient to increase wildfire potential, and future wildfire regimes are likely to be driven by temperature rather than rainfall132,133. Rising mean and extreme temperatures134, and increasing atmospheric aridity135, are linked to increases in fuel dryness, resulting in more frequent and severe wildfires4,9,31,49,136. Future projections of wildfire are region-dependent and subject to substantial uncertainties. But, in general, models point toward an increase in burned area for wildfire-prone regions such as western North America, southern Australia and the Mediterranean3,4,132,137, emphasising the continuing requirement for adaptation to changing wildfire patterns11.

Methods

Fire weather, drought and burned area

The McArthur Forest Fire Danger Index35,138 (FFDI) is calculated as

$${{{\rm{FFDI}}}}={D}^{0.987}\exp (0.0338T-0.0345H+0.0234W+0.243147),$$
(1)

where T (oC) is the maximum daily 2m temperature, H (%) is the daily average 2m relative humidity, W (km/h) is the maximum daily 10m wind speed. The “drought factor”, D, is the rolling 20-day total precipitation, scaled to range between zero and 10, with larger D for lower precipitation totals. We use the Japanese 55-year reanalysis product139 (JRA55), which provides data from 1958 to 2020 at a spatial resolution of 1.25o latitude and longitude. Very similar formulations of the FFDI have been used in studies of fire weather in Australia34,45. We chose to use the FFDI over indices from other major fire danger rating systems as it is simpler to compute and interpret. The Canadian Fire Weather Index is more widely used than the FFDI, but has more stringent requirements for its calculation36. For example, the index considers the possible effect of snowmelt affecting wildfire potential. Given that our study is focused primarily on the meteorological aspects of fire weather, we opted to use the FFDI.

We find that FFDI data in tropical regions is unstable in the early part of the record. Extremely high FFDI values, driven by very high T and W, and very low H, are present at various times for a large proportion of tropical grid cells before 1970. These values are often by far the greatest in magnitude throughout the entire record. We therefore exclude data for tropical grid cells (between the Tropic of Cancer, 23.44oN, and the Tropic of Capricorn, 23.44oS) before 1970.

To quantify fire weather over a given period, L, we accumulate the number of days on which the FFDI is above the climatological 95th percentile, calculated using all available data. Such days are called “fire weather days” (FWDs). On an annual basis, we accumulate fire weather days over a 12-month period that begins six months prior to, and ends five months after, the peak fire weather month. The peak fire weather month is determined as the month that yields the average maximum FFDI, as shown in Fig. 1d. Annual accumulations of fire weather days are denoted FWDf, where L = f indicates the “fire year”. Regionally-averaged quantities are denoted with a superscript indicating the region, r, i.e. \({{{{\rm{FWD}}}}}_{f}^{r}\). For monthly analyses, fire weather days are accumulated on a right-labelled rolling three-month basis, denoted \({{{{\rm{FWD}}}}}_{3}^{r}\).

We use satellite-derived monthly burned area data from FireCCI v5.1, provided by the European Space Agency Climate Change Initiative140 and C3S v1.0 from the Copernicus Climate Change Service. These data are provided on a 0.25o grid between 2001 and April 2020. Burned area is accumulated over different periods in the same way as for fire weather days: BAf and \({{{{\rm{BA}}}}}_{3}^{r}\) are defined using the same periods as their fire weather day equivalent.

We quantify drought using monthly precipitation from the Global Precipitation Climatology Centre (GPCC) dataset141 for 1957–2020. These data are derived from quality-controlled station gauge data, and are provided over land with a resolution of 0.25o latitude and longitude. As with fire weather days and burned area, we consider annual and monthly statistics. We estimate annual drought severity at the height of the fire season by accumulating precipitation in the 12 months leading up to the peak FFDI month. This quantity is denoted Pa, where the period L = a indicates the antecedent conditions. For monthly analyses of regional averages, right-labelled rolling 12-month accumulations are used, denoted \({P}_{12}^{r}\).

Throughout the manuscript we apply a mask to remove grid cells with zero total burned area (Fig. 1a). Data sets with coarser resolution than the burned area data are first aggregated to the required resolution by summing all burned area grid cells within each coarser grid cell.

Compound event definitions

We use FWDf, \({{{{\rm{FWD}}}}}_{f}^{r}\), Pa and \({P}_{a}^{r}\) to analyse annual changes in fire weather and drought. A “fire weather year” is defined as when \({{{{\rm{FWD}}}}}_{f} \,>\, {\mu }_{{{{{\rm{FWD}}}}}_{f}}+{\sigma }_{{{{{\rm{FWD}}}}}_{f}}\). Similarly a “drought year” is when \({P}_{a} \,<\, {\mu }_{{P}_{a}}-{\sigma }_{{P}_{a}}\). We deem a preconditioned drought and fire weather year to occur when both these conditions are satisfied. Superscripts are used in the usual way for the regional equivalents.

For the monthly analysis of fire weather in eastern Australia (EAU) and western North America (WNA), we define events in a similar way. A “fire weather (burned area) event” is a month in which \({{{{\rm{FWD}}}}}_{3}^{r}\) (\({{{{\rm{BA}}}}}_{3}^{r}\)) exceeds one standard deviation above the mean, i.e. \({{{{\rm{FWD}}}}}_{3}^{r} \,>\, {\mu }_{{{{{\rm{FWD}}}}}_{3}^{r}}+{\sigma }_{{{{{\rm{FWD}}}}}_{3}^{r}}\) or \({{{{\rm{BA}}}}}_{3}^{r} \,>\, {\mu }_{{{{{\rm{BA}}}}}_{3}^{r}}+{\sigma }_{{{{{\rm{BA}}}}}_{3}^{r}}\). A “drought event” is a month in which \({P}_{12}^{r}\) is less that one standard deviation below the mean, \({P}_{12}^{r} \,<\, {\mu }_{{P}_{12}^{r}}-{\sigma }_{{P}_{12}^{r}}\).

Trend and change point testing

The Mann-Kendall trend test142,143 is a rank correlation test between the ranks of the observations and their temporal ordering. For a time series X = {X1, X2, … , XT}, the test statistic, S, is given by:

$$S=\mathop{\sum }\limits_{i=1}^{T-1}\mathop{\sum }\limits_{j=i+1}^{T}{{{\rm{sgn}}}}({X}_{j}-{X}_{i}),$$
(2)

where

$${{{\rm{sgn}}}}({X}_{j}-{X}_{i})=\left\{\begin{array}{ll}1,\quad &{x}_{i} \,<\, {x}_{j}\\ 0,\quad &{x}_{i}={x}_{j}\\ -1,\quad &{x}_{i} \,>\, {x}_{j}.\end{array}\right.$$
(3)

The Pettitt test144 is a non-parametric used to identify possible step changes in a time series. The series X is said to have a change point at τ if Xt for t = 1, … , τ have a common distribution function F1(x) and Xt for t = τ + 1, … , T have a common distribution function F2(x), and F1(x) ≠ F2(x). The null hypothesis of “no change”, H0: τ = T, is tested against the alternative hypothesis of “change”, H1: 1 ≤ τ < T, using the statistic

$${K}_{T}=\mathop{\max }\limits_{1\le t < T}| {U}_{t,T}| ,$$
(4)

where

$${U}_{t,T}=\mathop{\sum }\limits_{i=1}^{t}\mathop{\sum }\limits_{j=i+1}^{T}{{{\rm{sgn}}}}({X}_{i}-{X}_{j}).$$
(5)

The Mann-Kendall and Pettitt tests assume the data are independent and identically distributed. This assumption is often not satisfied for meteorological data, which can exhibit serial dependence. We apply a block bootstrap procedure145,146,147,148 to try and mitigate the effects of autocorrelation. The block bootstrap randomly resamples “blocks” of length B. We assume that the data follow a first-order autoregressive process147. Then B is calculated as

$$B={(n-B+1)}^{(2/3)(1-n^{\prime} /n)},$$
(6)

where \(n^{\prime}\) is the effective sample size given by

$$n^{\prime} \approx n\frac{1-{\rho }_{1}}{1+{\rho }_{1}},$$
(7)

and ρ1 is the lag-1 autocorrelation coefficient. This procedure is applied to each grid cell, so B varies spatially. In our case B ranges between 1 and 11 for both Pa and FWDf (Supplementary Fig. 2).

We generate 1000 block-bootstrapped series of the same length as the original series and calculate the quantile q at which the test statistic (S or KT) calculated from the original data lies in the bootstrapped sample. The p value is then p = 1 − q.

To account for spatial autocorrelation and multiple testing, we control the false discovery rate149 (FDR). The FDR approach ranks the p values from the N tests in ascending order, p(1), …p(N). A test is considered statistically significant if p < pFDR, where

$${p}_{{{{\rm{FDR}}}}}=\mathop{\max }\limits_{j=1,\ldots ,N}\left\{{p}_{(j)}:{p}_{(j)}\le \frac{j}{N}{\alpha }_{{{{\rm{FDR}}}}}\right\}.$$
(8)

We choose αFDR = 0.1, allowing, on average, for 10% of statistically significant results to be erroneous (i.e. the null hypotheses are true).

Alongside Mann-Kendall statistical significance results, in Fig. 4 we show the magnitude of the trend estimated using the Theil-Sen estimator150. The Theil-Sen estimator is the median of all slopes between paired values. For the Pettitt test, we present the year of a change point (regardless of statistical significance) and the magnitude, computed as the difference in median FWDf between the period after, and prior to, the change point year.

Climate diagnostics

We obtain monthly sea-surface temperature (SST) data for 1958–2020 from the UK Met Office Hadley Centre Sea Ice and Sea Surface Temperature data set151. Atmospheric diagnostics are from JRA55. We use 500 hPa geopotential heights (z500) as an indicator of the mid-tropospheric circulation, zonal and meridional wind at 300 hPa (u300 and v300) as diagnostics of the upper-level circulation, and velocity potential at 200 hPa (ϕ200) as a measure of upper-level divergence and convergence. All data are aggregated to monthly means. We obtain anomalies by removing the monthly mean calculated over 1958–2020.

In Fig. 5, fire weather day and precipitation anomalies are accumulated over the periods f and a, respectively, as described earlier. As such, this period varies spatially. For SST and the atmospheric variables, however, we use a common period for all grid cells to avoid averaging over a number of different periods. We choose the period based on the most common peak FFDI month for the Northern Hemisphere (June), the tropics (August) and the Southern Hemisphere (January). Twelve-month anomalies (left column) are averaged from six months before, to five months after, the listed months. Three-month anomalies (right column) are centred on the listed months.

We calculate indices for five climate modes. To represent El Ni\(\tilde{{{{\rm{n}}}}}\)o Southern Oscillation (ENSO), we use the Ni\(\tilde{{{{\rm{n}}}}}\)o3.4 index152, which is the average SST anomaly over 5oN-5oS, 120o-170oW. The Indian Ocean Dipole (IOD) is represented using the Dipole Mode Index153 (DMI). The DMI is the difference between SST anomalies averaged over western (10oN-10oS, 50o-70oE) and south-eastern (0o-10oS, 90o-110oE) regions in the Indian Ocean. The Southern Annular Mode (SAM) is represented using an index defined as the difference between the normalised mean sea-level pressure at 40oS and 60oS113. The Pacific-North America pattern (PNA) index126 is calculated as PNA = (z1 − z2 + z3 − z4)/4, where z1, z2, z3 and z4 are the z500 anomalies at 20oN, 160oW; 45oN, 165oW; 55oN, 115oW; and 30oN, 85oW; respectively. The Gulf of Alaska ridge (GAR) index130,131 is calculated as 200 hPa geopotential height anomalies averaged over the domain 45o-55oN, 145o-135oW.

In Fig. 6, these indices are presented as three-month averages. The months were chosen as those that yield the greatest correlation with December \({{{{\rm{FWD}}}}}_{3}^{{{{\rm{EAU}}}}}\) (for Ni\(\tilde{{{{\rm{n}}}}}\)o3.4, DMI and SAM) or August \({{{{\rm{FWD}}}}}_{3}^{{{{\rm{WNA}}}}}\) (for GAR). As the PNA is typically considered a winter phenomenon126,127,128,129, and winter values of the index have been related to wildfire by other studies92,93, December through February was chosen as the appropriate period for this index.