Drought triggers and sustains overnight �res in North America

: Overnight fires are emerging in North America with yet unknown drivers and implications. This phenomenon is significant as it challenges the traditional understanding of the “active day, quiet night” paradigm of the diurnal fire cycle and current fire management practices . Here we show that ~20% of all large fires (≥1000ha) in North America observed during 2017-2020 using geostationary satellite images and terrestrial fire records burned through a total of 1,084 nights. Overnight burning was characterized by early onset after ignition, high persistence, and a tendency to mutually reinforce extreme fires. While warming weakens the climatological barrier to general nighttime fires 1 , we found the major driver of overnight burning was the accumulated fuel dryness and fuel availability (i.e., drought conditions), rather than fast -reacting day-night weather fluctuations. Drought conditions tended to sustain overnight burning for periods of multiple days, and even weeks. Moreover, we show that daytime drought indicators can be used to predict overnight burning events, which could facilitate early detection and management of nighttime fires. Recently observed and predicted future increasing trends in conditions conducive to overnight burning indicate that disruption of the diurnal fire cycle may accelerate, leading to larger, more intense, and extreme fires in future.


Main Text
Asymmetric warming, where nights are warming more rapidly than days resulting from anthropogenic climate change 2,3 , may significantly impact diurnal fire activity.Although changing daytime conditions 4 are known to exacerbate fires 1 , the potential shifts in nighttime burning have received less attention, as nighttime fires are typically hindered by cooler and moister atmospheric conditions and increased moisture in the fine fuels 5 .This conventional understanding of day-night fire pattern has been widely applied to fire suppression 6 , and prescribed 7 , cultural 8 , and agricultural 9,10 burning.However, recent reports from frontline firefighters and satellite observations have indicated an increase in the frequency and duration of nighttime fire incidents in Canada and the USA 11 and an increase in the number and intensity of nighttime fire "hotspots" 1,12 .These findings raise concerns that increasing nighttime flammability in certain regions may be pushing the general diurnal burning pattern towards a tipping point, where the absence of nighttime conditions acting as a break to fire activity could lead to overnight burning self-perpetuating and thus resulting in more extreme fire events.Satellite-based active fire products provide spatiotemporal observations of fire activity 13 , however, existing research [14][15][16] on fire diurnal cycles has primarily focused on regional scale patterns, with hotspots and fire intensity often peaking in the mid-afternoon.Studies have shown that nighttime fire seasons are shorter than daytime fire seasons in most Australian 5 and global 1 climate regions.Nighttime hotspots tend to be present during the peak fire season 17 and are more commonly associated with large fires, especially in arid regions 18 and under extreme droughts 12 .While local cases of overnight fire have been documented 19 , we were unable to find any methodical studies exploring this urgent phenomenon and its implications, or its underlying driving factors.
Fire activity has been widely linked to weather conditions captured by fire weather indices and meteorological parameters, such as the Canadian Fire Weather Index System (CFWIS) 20 and vapor pressure deficit (VPD, a widely used metric measuring how rapidly the atmosphere dries fuel) 21,22 .CFWIS components are the most commonly used indices for both operational and research purposes regionally 23,24 and globally [25][26][27] .CFWIS firstly tracks potential fuel moisture conditions in surface fine fuel and moderate and deep organic layers at daily or hourly timesteps 28 , capturing the varying speeds with which these fuels react to ambient weather.Using the resulting fuel moisture codes, CFWIS then derives indices of potential fire behavior: potential rate of spread, available fuel, and fire intensity 29 .The extent to which fires can burn at night is partly governed by diurnal weather fluctuations 30 and corresponding changes in surface fuel moisture 31 .The day-night extreme values and range of these factors can be important for overnight fires as they largely determine both the initial conditions at the start of the night, and nighttime minima, while a smaller day-night range may sustain longer-lasting nighttime burning.Dry fuels that react slowly to meteorological conditions may also play a key role in overnight burning as they provide relatively stable influences on fires 12,32 despite diurnal meteorological fluctuations.While recent research has examined the instantaneous relationship between nighttime fire activity and hourly VPD 1 , the systematic effect of slow-and fastreacting fuels and weather dynamics on overnight burning remains unexplored.
Here we used active fire detection data from the Geostationary Operational Environmental Satellite-R Series 33 combined with high-accuracy terrestrial fire datasets to characterize the hourly fire diurnal cycle of each large fire (≥1000 Ha) and identify overnight burning events (OBEs, i.e., fire activity was detected during every nighttime hour in a single day within a known fire perimeter) in North America during the 2017-2020 period.The distribution, characteristics, and implications of OBEs were further analyzed.A systematic examination of fire weather metrics (including CFWIS and VPD) based on European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5) 34 , was conducted to assess underlying drivers and explore the prediction of OBEs.To our knowledge, this is the first effort to comprehensively explore key aspects of overnight fires and could greatly fill the knowledge gap of diurnal fire activity and its changing nature, and practically contribute to nighttime fire management.

Overnight burning is widespread and extreme
We studied the diurnal cycles of 1,655 large fires that occurred in North America during 2017-2020 (Supplementary Fig. S1) and found that 20% (n=330) of these fires experienced overnight burning (Fig. 1 and Supplementary Table S1).A total of 1,084 OBEs was identified, which accounted for 2.2% of the total nights of all fires.Of these OBEs, 86.2% were found in the dominant multi-OBE fires (n=180).OBEs were most concentrated in the Boreal region (13.4%) and western mountain areas (Temperate mountain system: 45%, and Subtropical mountain system: 30.4%),where large fires were also the most prevalent.The proportions of fires with OBEs in temperate and subtropical mountain systems were as high as 35.2% and 35%, respectively, and about two-thirds (65.1% and 63.5%, respectively) of them were multi-OBE fires, which experienced 5.4 and 6.4 OBEs on average.OBEs peaked in summer (June-August) and fall (September-November) in western mountain areas with few in spring (March-May), while around a quarter of OBEs in the Boreal occurred in spring, especially in Alberta (Fig. 1).Fires with and without OBEs are represented by black and gray triangles, respectively.The color and size of filled circles represent the season (e.g., summer: June-August) and the number of OBEs per fire, respectively.The background is colored by biome classifications.When OBEs span multiple seasons, the geographic position is jittered and plotted multiple times with different colors for clearer visualization.The numbered green triangles represent two extreme fire cases that are discussed further in the Supplementary Text: (1) the 2020 Creek Fire (California, Subtropical mountain system); and (2) the 2019 McMillan Complex Fire (Alberta, Boreal).The definition and identification approach for OBEs can be found in Methods.
We observed a strong association between OBEs and extreme fires, providing evidence for the existence of a mutually reinforcing relationship between them.More than half of OBEs occurred in just 13% of the most extreme wildfires (Fig. 2a), with the top 10 fires having an average of 27.1 overnight burning nights (Fig. 1).Analysis of the number of days between two consecutive OBEs indicated most OBEs in multi-OBE fires are temporally clustered, often occurring continuously or within a short time interval during the lifetime of a fire (Fig. 2b).For example, 43 OBEs were identified within 51 days of the Creek Fire in California during the Submitted Manuscript: Drought triggers and sustains overnight fires in North America 6 2020 fire season (Fig. 1 and Supplementary Fig. S5).The average size of fires with OBEs (31,071 ha) was 6.6 times greater than fires without OBEs (4,682 ha), and there was a positive relationship between fire size and the number of OBEs (Fig. 2c).Furthermore, OBEs tend to occur in the first few days after ignition, with more than half of fires with OBEs experiencing overnight burning within 2 days of ignition, and nearly one-third on the day of ignition (Fig. 2d).Notably, the number and characteristics of OBEs reported herein are likely to be conservatively estimated, given our stringent requirements for identifying an OBE (Methods) and the omission errors of Earth observation-based active fire detection algorithms (e.g., obscuration, oblique sensor observation angles, and smoldering fires) 35 .

Fig. 2. The extreme characteristics of OBEs.
(a) Cumulative percentage of fire with OBEs vs. cumulative percentage of OBEs by biome and for all biomes combined ("All").Red dashed lines indicate that 13% of fires with OBEs contributed to more than half of all OBEs.(b) The frequency and distribution of the number of days between two successive OBEs in multi-OBE fires.In 67% of cases, fires burned continuously through two consecutive nights, indicating a high persistence of OBEs.(c) Fire size vs. the number of OBEs per fire by biome and for all biomes combined ("All").Relationships are fitted using linear regression, and in all cases, P<0.05, indicating that fire size is proportional to the number of OBEs.(d) The frequency and distribution of the number of days between fire ignition and the occurrence of the first OBE, for all fires with OBEs.52% of these fires experienced an OBE within 2 days of ignition and 31% experienced an OBE on the day of ignition.

Drought drives overnight burning
We examined the differences in all fire weather variables (i.e., daily slow-reacting variables and daytime and nighttime extremes of hourly fast-reacting variables; Methods) and the daynight range of fast-reacting variables between OBEs and non-OBEs in the five major biomeseason groups with 100 or more OBEs (Boreal summer, Temperate mountain system summer and fall, and Subtropical mountain system summer and fall; Supplementary Table S1).
However, the day-night ranges of all fast-reacting weather variables for OBEs were not significantly smaller than for non-OBEs (Supplementary Fig. S3).Moreover, the value of each fire weather variable during OBEs generally exceeded the 90 th percentile of comparable observations during 2000-2020 in the same location (Fig. 3c and Supplementary Fig. S4), for example, BUI and DMC exceeded the 93 rd and 92 nd percentiles, respectively.Notably, fire weather percentile values during OBEs were significantly lower (paired Wilcoxon test) when compared to a 2000-2020 co-located fire weather climatology than when compared to a 1979-1999 climatology (Fig. 3c), indicating an increasing trend in extreme day-night fire weather conducive to overnight burning in recent decades.To investigate the major drivers of recent OBEs, we constructed random forest binary classification (i.e., OBEs and non-OBEs) models 36 for the five major biome-season groups to determine the relative importance of fire weather variables (Methods).The results indicated that drought-related variables played a crucial role in supporting OBEs (Fig. 4).Specifically, either DMC or BUI was found to be the most important factor in all groups.In contrast, fastreacting weather variables were relatively less important e.g.temperature (T) and relative humidity (RH) based variables were generally low-ranked in all biome-season groups.
Although surface fuel moisture (FFMC) and potential fire spread (ISI) were not as important as BUI and DMC overall, they strongly influenced OBEs in Boreal summer and Temperate mountain system summer.Moreover, the qualitative analysis of the 2020 Creek Fire in California (Supplementary Fig. S5 and Text) also illustrated the dominant role of drought conditions on OBEs.For each major biome-season group, we calculate the normalized mean decreases in the Gini coefficient of fire weather variables based on the random forest model for classifying OBEs and non-OBEs.The variables are ranked from the most important to the least important.Slow-reacting variables are represented by dark red horizontal bars, and daytime and nighttime extremes of fast-reacting variables by gray and black bars, respectively.The performance of the models is evaluated by the area under the receiver operating characteristic curve (AUC).

Overnight burning is predictable
In operational wildfire management, fire danger indices and adjective ratings used for decisionmaking are typically generated at a daily (i.e., local noon) timestep, rather than an hourly one, especially in remote areas 37 .To explore the potential predictability of OBEs in such an operational setting and to understand the relationship between daytime conditions and nighttime burning, we constructed logistic regression models for the five major biome-season groups using different combinations of daily fire weather variables (Fig. 4).The results indicated that OBEs were predictable and that daytime conditions largely set the foundation for their occurrence.For each biome-season group, at least 70% of OBEs were correctly predicted (Supplementary Table S2).For example, in the best-performing model for the Temperate mountain system fall, 82.6% of OBEs were correctly predicted (Fig. 5).

Overnight fires: an understudied but emerging challenge
Compared to interannual, annual, and seasonal fire activity, diurnal fire activity, especially the nighttime aspect, has long been overlooked.However, the recent widespread occurrence of unexpected and extreme OBEs in conventional large-fire-prone areas in North America has highlighted the urgency of this research.These OBEs are characterized by early onset after ignition and high persistence, and are found to be closely associated with extreme fires, challenging traditional diurnal fire knowledge and current fire management practices.The primary cause of OBEs is extreme fire weather, particularly intensified fuel dryness and availability, which disrupts the usual balance of diurnal flammability.Our study also emphasizes the predictability of OBEs, providing new insights into the diurnal fire cycle with implications for nighttime fire management.
We identified two main drivers of OBEs: DMC and BUI, which aligns with previous studies that have suggested that nighttime fires (e.g., higher occurrence 18 and longer persistence 12,19 ) favor drier conditions.These two drivers not only react slowly to diurnal fluctuations but also exhibit time-lags of days to weeks 38 , which prevent fires from being extinguished during adverse nighttime conditions.This may also explain why OBEs usually occur on consecutive or nearly consecutive nights (Fig. 2b).Nonetheless, the patterns of OBEs may still vary between regions and seasons.For instance, nearly a quarter of Boreal OBEs were identified in spring when fuel dryness and availability usually cannot accumulate sufficiently.A qualitative analysis of the 2019 McMillan Complex fire, Alberta (Supplementary Fig. S6 and Text) and previous research on large spring fires in Alberta 39 indicate that spring OBEs in the Boreal are likely wind-driven.Additionally, sudden changes in nighttime conditions, such as the passage of a dry cold front 40 or the onset of heatwave conditions, can also weaken or even eliminate the nighttime barrier to fires, resulting in OBEs.
Overnight burning presents significant challenges for fire management.Firstly, conditions conducive to OBEs typically occur when fire suppression capacity is already stretched.The extended burning duration, larger burned area and intensity, and extreme fire behavior can exponentially increase containment expenses 41 .Secondly, the early onset of OBEs after ignition leaves little time for firefighters to react, and the persistence of OBEs limits containment options.Multi-OBE fires, the dominant form of overnight burning, are therefore harder to extinguish and more likely to become escaped fires.Thirdly, firefighters face limited time for rehydration, sleep, and reducing body temperature, exacerbating physical and mental stress 42 .Reduced visibility and more complicated nighttime situations further escalate this adversity 12,43 .To cope with these challenges, early fire detection efforts and developing new tools that allow for more effective decision-making may be viable approaches given the increasing budget pressures on fire management 44 .For example, as we show here for the first time, OBEs are predictable based on daytime conditions.Combining these findings with fire weather forecasting 45 and real-time data assimilation of observations in an operational system could enhance strategic and tactical management decisions.
Recent decades have seen a rise in day-night extreme conditions conducive to OBEs, which is consistent with the prolonged drying period (e.g., extreme droughts in the western USA) 46,47 and increasing trends in fire-conducive weather during the day or night 1,2,48 .However, the relationship between diurnal fire activity and climate change remains largely understudied, particularly regarding the non-linear impact of asymmetric warming, where a smaller increase in daytime temperature may have a disproportionately greater impact on diurnal flammability.
To better understand this relationship, future research should involve not only climatological factors 1 but also fuel factors and evaluate it in a day-night integrated manner, rather than separately, given the dominant drought-driven pattern of OBEs and the association between daytime conditions and OBEs.Moreover, insights into how diurnal burning patterns are expected to shift regionally and globally in the future can provide both scientific and practical values for confronting future fire challenges 49,50 .
States and ≥500 acres (~202 Ha) in the eastern United States and provides attributes such as ignition date and area burned 54 .
Geostationary active fire detections.The Fire/Hot Spot Characterization Full Disk (FDCF) products from both GOES-16 (May 2017 to present) and GOES-17 (August 2018 to present) were used to obtain the sub-hourly active fire detections ("hotspots") in North America during the period 2017-2020.The FDCF products were downloaded from their first available dates to 2020 in Amazon Web Service S3 Explorer (https://registry.opendata.aws/noaa-goes/).These products use both visible and infrared Advanced Baseline Imager (ABI) spectral bands to locate fires and retrieve sub-pixel fire characterizations with 5-15 minute temporal resolution and a nominal 2-km spatial resolution (coarser with increasing distance from the sub-satellite point) 33 .These satellites are positioned at 75.2° W and 137.2°W, respectively, and can observe the entire burnable land in North and South America when used together.The nighttime active fire detection algorithm is considered to be more sensitive to smaller and/or cooler fires than the daytime algorithm because ambient background temperatures are more homogeneous and lower at night, increasing the potential contrast provided by active fire pixels 35 .
Notably, the availability, frequency, and quality of the fire detection data varied both regionally and over the course of the study period, as a result of the sequential launch of GOES-16 and GOES-17, changes in imaging frequency, and the coverage area and view zenith angle of each instrument (for more detail on these variations, please see Supplementary Text).Due to these influences, we restrict our use of the hotspot dataset here to identifying the hourly burning status of the individual fires recorded in the wildland fire databases.The available hotspot data had a minimum observation frequency of four times per hour at any location within the study area, and as such, we considered these data fit for this purpose.
Hourly fire diurnal cycle and OBEs identification.We developed an algorithm to characterize the hourly fire diurnal cycle of large fires (≥1000 Ha) in two datasets (NBAC and MTBS) and identify OBEs (Supplementary Fig. S1).The start (end) date for fires in NBAC was the earlier (later) date between the detected start (end) date and the reported start (end) date.For fires in MTBS, the start date was the recorded ignition date.As the MTBS does not contain information on fire end dates, end dates were inferred from the GOES-R hotspot data based on the last date on which two consecutively active burning hours were determined.The same approach was used to infer start and end dates for NBAC fires with missing start (1/472) and end (2/472) date records.GOES-R hotspots intersecting fire perimeters from the fire start and end dates were extracted.Fires without matching hotspots were excluded from further analysis.Fires missing corresponding hotspots data may be due to (1) no available active fire data in certain regions and/or periods of the study (Supplementary Text) or (2) active fire detection algorithm omission errors 35 .This resulted in a total of 1,655 out of 2,390 fires being studied, with 472 fires from NBAC and 1,183 from MTBS.A fire was considered active in a specific hour when at least one hotspot was detected within its perimeter during that hour.
The resulting hourly fire activity data were converted from Universal Time Coordinated (UTC) time zone to local time zones based on the spatial centroid of the fire perimeters and Day-Of-Year (DOY).The exact times of sunrise and sunset were used to separate daytime and nighttime for each day (i.e., a 24-hour period spanning the first hour after sunset to the last hour before sunrise of the next day), with sunrise referring to the top edge of the sun appearing on the horizon, and sunset referring to the top edge of the sun disappearing below the horizon.Using these local times, each hour was designated as either a daytime or nighttime hour and further assigned to one of four categories: active daytime, non-active daytime, active nighttime, or non-active nighttime.For all fires where the number of nighttime hours was ≥4 for each day, an OBE was identified when hotspots were detected in each nighttime hour.Days where fire activity did not occur in every nighttime hour were classified as non-OBEs.A 4-hour threshold was applied to exclude part of days of high-latitude summer fires from further analysis.This is because the extent of weather and fine fuel moisture fluctuations is limited during very short nighttime periods 55,56 , making it unsuitable for studying the impact of changes in nighttime conditions on overnight burning.Time zone conversions and local sunrise and sunset times were obtained through the use of R packages lutz and suncalc.
Fire weather calculation and extraction.In this study, meteorological variables provided by ERA5 reanalysis data were used to process and calculate fire weather from 1979 to 2020, including the inputs and components of CFWIS as well as VPD.ERA5 is the fifth-generation of global hourly atmospheric reanalysis produced by the European Center for Medium-Range Weather Forecasts (ECMWF) and is widely used in wildfire studies 4,57 .It resolves the atmosphere using 137 levels from the surface up to a height of 80 km on a 31-km horizontal grid 4 .
The CFWIS usually outputs six daily components by firstly tracking moisture in three fuel layers of varying depth with corresponding moisture codes: Fine Fuel Moisture Code (FFMC, litter and fine fuels), Duff Moisture Code (DMC, organic fuels at moderate depth), and Drought Code (DC, deep and compact organic fuels).The remaining three components are potential fire behavior indices: Initial Spread Index (ISI, the rate of fire spread), Buildup Index (BUI, the cumulative fuel availability), and Fire Weather Index (FWI, the fire intensity).To account for the impact of diurnal fluctuations in weather and surface fuel moisture on OBEs, the hourly FFMC and ISI were calculated using the procedures outlined in ref. 56,58 .The hourly FFMC calculation was based on the hourly weather observations of 2-m temperature (T, °C), relative humidity (RH), wind speed (WS, km h -1 ), and precipitation, as well as the previous hour's weather conditions.RH was calculated from 2-m T and 2-m dewpoint T following equations 1 and 2 in ref. 59 .WS was calculated from 10-m U (zonal velocity) and V (meridional velocity).
The hourly ISI combined hourly FFMC and hourly WS.The hourly VPD was calculated based on the conversion equation from ref. 60 using 2-m T and 2-m dewpoint T. The remaining four components of the CFWIS (FWI, DMC, DC, and BUI) were obtained from ref. 4 at daily temporal resolution as they or their subcomponents are slow-reacting to weather fluctuations.
We extracted the aforementioned fire weather variables for each large fire during its lifetime, buffering 24 h at the start date, and then time-matched these data with the corresponding fire diurnal cycles.For each timestep, data were spatially averaged across all grid cells intersected by a given fire perimeter.
The spatiotemporal distribution and statistics of OBEs.We summarized the and temporal distribution of OBEs by biome and season.Fires with OBEs were further categorized into single-OBE fires and multi-OBE fires, where only one OBE and more than one OBE occurred, respectively.The proportion of multi-OBE fires to all with OBEs and the mean number of OBEs per multi-OBE fire was calculated by biome.
Extreme characteristics of OBEs.For all fires with OBEs, we calculated the number of days between ignition and the occurrence of the first OBE, and the number of days between two adjacent OBEs in multi-OBE fires to evaluate the persistence of OBEs and the potential impact of OBEs on fire management.Moreover, fire size was compared between fires with and without OBEs, and linear regression (significance level: 0.05) was used to investigate the relationship between the number of OBEs and fire size for fires with OBEs in all biomes as well as separately in Boreal, Temperate mountain system, and Subtropical mountain system (three major biomes with most OBEs).
Comparison of fire weather conditions.We examined the differences in the fire weather between OBEs and non-OBEs (including non-OBEs during fires with OBEs) in the five major biome-season groups with 100 or more OBEs (Boreal summer, Temperate mountain system summer and fall, and Subtropical mountain system summer and fall), using a one-sided Mann-Whitney U test (significance level: 0.05) to investigate significant drivers of OBEs.We also evaluated differences in the day-night range of hourly variables between OBEs and non-OBEs using a one-sided Mann-Whitney U test, as we hypothesized that OBEs may experience a smaller diurnal range in fire weather which may facilitate their occurrence.These ranges were FFMCDmax-Nmin, ISIDmax-Nmin, RHDmin-Nmax, TDmax-Nmin, and VPDDmax-Nmin.The Mann-Whitney U test was selected as the distribution of some variables was skewed.
Assessment of and increasing trend in fire weather extremes of OBEs.To assess the fire weather extremes during OBEs, we calculated the percentile value of each OBE's fire weather variable relative to the distribution of values extracted from records for the years 2000-2020 and 1979-1999 in the corresponding fire perimeter.The fire weather variables examined were the four daily components (FWI, BUI, DMC, DC) of CFWIS and, based on hourly variables, the daytime extremes (FFMCDmax, ISIDmax, VPDDmax, TDmax, and RHDmin) and nighttime extremes (FFMCNmin, ISINmin, VPDNmin, TNmin, and RHNmax), capturing the effects of slow-and fast-reacting fuel and weather dynamics.We then used a paired Wilcoxon test to compare these two percentiles for each OBE's fire weather variable to determine if there has been a change in diurnal fire weather between the two time periods.
Importance analysis of fire weather variables.We aimed to understand the dominant factors influencing OBEs by building random forest (RF) models 36 to analyze the relationship between fire weather and OBE occurrence for the five major biome-season groups (see 'Comparison of fire weather conditions').The RF models were built using a binary dependent variable (OBE or non-OBE) and all fire weather variables as the explanatory variables.RF is an ensemble approach that consists of many individual decision trees, which take input from randomly bootstrapped variables and samples.In RF ensembles, we chose 500 trees within each forest and a depth of 1 for each tree.To ensure statistically reliable results, we repeated the process 50 times using 5-fold cross-validation.To resolve the impact of the disparity in frequencies of the two classes, non-OBEs data were down-sampled in each training.We used the mean decrease in the Gini coefficient (i.e., Gini importance) to measure how each variable contributes to the impurity decreases of the tree nodes in the resulting RF.The higher the value of the Gini importance, the higher the importance of the variable in the model.We normalized the Gini importance values and used the area under receiver operator characterization (ROC) curve (AUC) 61 to evaluate the performance of models.Model construction and performance evaluation were performed using the R package caret 62 .Two extreme fire case studies were selected to demonstrate the relation between the occurrence and persistence of OBEs and fire weather varying in time and space: the 2020 Creek Fire (California, Subtropical mountain system) and the 2019 McMillan Complex Fire (Alberta, Boreal) (Supplementary Text).
Prediction of OBEs.Logistic regression (LR) prediction models were built for the five major biome-season groups (see 'Comparison of fire weather conditions') using either a single daily variable (4 models) or a combination of multiple daily variables (11 models).We discarded models that incorporated multiple variables if the variance inflation factor of any variable was greater than 2 to avoid multicollinearity.Similarly, 50 times 5-fold cross-validation and downsampling in each training were also performed for each prediction model.Model performance was evaluated using ROC, AUC, and recall (i.e., the proportion of OBEs that were successfully predicted by the model).The best models were determined by both AUC (primary criterion) and recall (secondary criterion).Model building and performance evaluation were conducted using the R package caret.

Fig. 3 .
Fig. 3. Fire weather is elevated during OBEs and has become more extreme over time.Significant greater (one-sided Mann-Whitney U test, P < 0.05) fire weather conditions during OBEs (red for summer and orange for fall) than those for non-OBEs (gray for summer and black for fall) in the (a) Boreal and (b) Temperate mountain system.We invert the y-axis of these distributions in fall for clearer visualization.Details for all variables in the five major biome-season groups are shown in Supplementary Fig. S2.(c) The line-linked paired points respectively represent the percentile of fire weather for each OBE relative to comparable observations during the 1979-1999 and 2000-2020 periods at the same geographic location.The 1979-1999 percentiles are significantly higher than the 2000-2020 percentiles for each fire weather variable (paired Wilcoxon test, P < 0.05).Box plots display the distribution of these percentile values, with a median line, mean triangle, and box ends representing first and third quartiles.Whiskers extend to values within 1.5 times the inter-quartile range.Details for other fire weather variables are shown in Supplementary Fig. S4.

Fig. 5 .
Fig. 5. Skill of OBE prediction models.Model for each major biome-season group is built using logistic regression.The receiver operator characterization curve (ROC) of each resample (background gray lines) for a 50 times 5-fold cross-validation and the ROC from all resamples (colored lines) are shown in each subplot, with the overall area under ROC (AUC) value presented.The recall represents the percentage of correctly predicted OBEs among observed OBEs.The equations in each subplot show the logistic equations for the model output in different biomes, where P represents the probability of OBE occurrence.The overall ROC, AUC, recall, and logistic equations are colored by season, i.e., red for summer and orange for fall.