A weakened diurnal weather constraint leads to longer burning hours in North America

Abstract

Contemporary North American wildfires exhibit increasingly erratic intraday burning, posing immediate operational and socioeconomic challenges. Here, we show that climate-driven weakening of day-night (diurnal) weather constraints extends and intensifies burning hours, a key mechanism behind broader fire regime transformations. Analyzing hourly geostationary satellite observations for ~9000 fires (>200 hectares; 2017–2023), we found western mountains and boreal forests experienced the longest active burning hours, with approximately one-third of active days exceeding 12 hours. About 60% of fires reached peak intensity within 24 hours of detection, while 14% of active days peaked at night. On the basis of fire weather, annual potential burning hours were estimated to rise 36% over 1975–2024, with pronounced increases in western regions and spring/fall (48 to 57%). Regions with significant changes gained 26 more potential active days annually and 1.2 additional potential burning hours daily, while extreme days (≥12 or 24 potential burning hours) rose 81 to 233% in fire-prone biomes. Future management requires adaptation to wildfires that increasingly defy diurnal norms.

Weakened constraints of the day-night weather cycles are making wildfires burn longer and harder to control across North America.

INTRODUCTION

Wildfire regimes across North America are undergoing a profound transformation, driven by a combination of changing climate (1–3), altered land use (4, 5), legacy fire suppression policies (6–8), and shifting ignition patterns (9, 10). Recent research has advanced our understanding of wildfire trends through cumulative or retrospective metrics—such as seasonal and annual burned area (3, 11), emissions (12), severity (13, 14), and fire season length (15, 16). However, these broader indicators fail to capture the increasingly erratic intraday dynamics that disrupt the traditional diurnal cycle (i.e., the 24-hour day-night cycle), reshaping modern fires and posing the most immediate operational and socioeconomic challenges, as increasing fires were able to spread up to tens of kilometers within hours and sustain weeks of overnight burning (17–20). These diurnal surges not only outpace traditional tactics (e.g., nighttime suppression) and resource mobilization but also place communities at acute risk with compressed evacuation timeframes and overwhelmed local defenses (21–23), as tragically demonstrated by the 2023 Maui Fire (Hawaii) (24), the 2024 Jasper Fire (Alberta) (25), and the 2025 Los Angeles Fires (California) (26).

Satellite-based active fire products have provided consistent, large-scale observations (27–29), allowing fire activity to be linked to fire weather conditions at varying temporal scales. However, fire pattern research largely relies on low-Earth orbit satellites, whose 12-hour or longer revisit intervals adequately capture regional-scale patterns (30) but fail to resolve critical hourly dynamics of individual fire events. Geostationary satellites [e.g., Geostationary Operational Environmental Satellite-R (GOES-R) series] can capture active fire data subhourly (29). Their main applications have been focused on early detection (31, 32) and smoke monitoring (33) or analyses of regional megafires (34) and nighttime behavior (17, 35).

Hourly fire dynamics are influenced by fast-reacting variables (e.g., diurnal changes in weather and fine fuel moisture) and slow-reacting conditions (e.g., moisture in large-diameter dead fuels and subsurface organic materials) (36–38). We expect that fast-reacting variables directly influence the hourly burning probability and intensity, and slow-reacting conditions shape the baseline fire susceptibility and amplify the effects of hourly fluctuations (17). Prolonged burning hours can overstretch firefighting resources and increase burning severity, accelerate emissions, and threaten ecosystem resilience (13, 17, 39–41) as they are associated with more extreme fire conducive conditions. For fire management agencies, accurately estimating burning hours each day is essential for designing operational tactics (42, 43), scheduling suppression crews, and running simulation tools (e.g., BurnP3) (44) that rely on realistic inputs to simulate fire ignition and growth. Although anecdotal evidence and increasing nighttime burning (17, 35, 43, 45) suggest shifts in intraday or hourly fire patterns, continental-scale quantitative evidence of their long-term change also remains understudied.

Here, we integrated high-temporal-resolution GOES-R satellite active fire observations with high-spatial-resolution fire perimeters to systematically analyze hourly fire dynamics across North America during 2017–2023. We examined how the duration and intensity of active burning hours (ABH; hours with ≥1 active fire detection) evolve throughout each day and over the entire fire event across diverse ecosystems. Leveraging machine-learning models built upon hourly climate variables and fire weather indices, we modeled hourly burning probability and reconstructed historical potential burning hours (PBH; hours with conditions modeled to support active burning) from 1975 to 2024. We quantified intraday, seasonal, and annual PBH shifts across ecosystems and subnational jurisdictions over this period, revealing long-term changes in hourly fire potential. Our results highlight the evolving nature of wildfire behavior, emphasizing the operational value of hour-by-hour metrics. This refined perspective is critical for enhancing suppression strategies, improving fire simulation models, and informing policy in an increasingly flammable North America.

RESULTS

Large fires burn with longer active hours

Overview of fire activity

We analyzed 8993 fires (>200 ha) across North America (2017–2023) using hourly GOES-R active fire detection, identifying ~350,000 ABH across ~50,000 active days (days with ≥1 ABH). A longitudinal gradient emerged in ABH and fire radiative power (FRP; a proxy of burning intensity), with western North America hosting the most extreme events (Fig. 1A and fig. S1). Three fire-prone biomes—temperate mountain systems (36% of ABH), boreal forests (32%), and subtropical mountain systems (12%)—dominated total ABH (Fig. 1A; biome classifications shown in fig. S2). While around half of active days in these biomes had <6 ABH, approximately one-third in subtropical and temperate mountains sustained ≥12 continuous ABH, with 12% of active days burning over a full 24-hour period in subtropical mountains (fig. S3). Seasonally, summer (June–August) showed longer daily ABH across most biomes, while western boreal biomes reached their peak daily ABH in spring (table S1).

Fig. 1. Observed characteristics of ABH and FRP for large North American wildfires, 2017–2023.

(A) Maps of fires (>200 ha) as gray dots (United States and Canada), with the top 200 fires with the most ABH from United States and Canada, respectively, in color; light gray outlines states, provinces, and territories. Cumulative hourly maximum FRP map in fig. S1. (B) Daily ABH versus daily maximum FRP (megawatt, MW); box plots show median (line), mean (red dot), quartiles (box ends), and whiskers (1.5× interquartile range); outliers >8000 MW were omitted for clarity; plot with all samples can be found in fig. S4. (C) Fire-event size versus maximum FRP versus total ABH, showing larger fires burn longer and more intensively. (D) Hours from first detection to fire-event maximum FRP; ~60% peak within 24 hours with ~17% being single-day fire events.

Hourly intensity metrics

We tracked two FRP-based intensity metrics per fire within its perimeter: daily maximum FRP (highest hourly FRP in a day) and fire-event maximum FRP (highest across the event), approximating fire front intensity critical to explosive and destructive spread (46). Prolonged ABH strongly correlated with higher maximum FRP at both daily (Fig. 1B and fig. S4) and fire-event scales (Spearman’s rank = 0.70, P < 0.01). Larger fire events tend to burn longer and more intensively (Fig. 1C and fig. S5). Notably, ~60% of fires reached their fire-event maximum FRP within 24 hours of first detection (Fig. 1D) with ~17% being single-day fire events (fig. S6), and 14% of active days with nighttime burning peaked in daily FRP at night, narrowing suppression windows.

Long-term trends: Rising hourly fire potential

Modeling hourly fire dynamics

We trained random forest (RF) models to estimate hourly burning status (active or not) on active days using the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5) (47) climate variables and the Canadian Fire Weather Index System (CFWIS) indices (48) (Materials and Methods). Evaluated under three independence schemes for train/test splitting (fire-event, fire-day, and fire-hour), the models showed good and robust discrimination, with AUC (area under the ROC curve) spanning 0.79 to 0.90 and sensitivity (recall of the active-burning-hour class) around 0.76 to 0.79 (Materials and Methods and table S2). Across interpretation diagnostics, the combination of all hourly fire weather variables accounts for the largest share of performance gains with relative humidity and temperature forming the first tier drivers, i.e., hotter, drier conditions increase burning probability (Materials and Methods and fig. S7). While topography and ignition patterns also influence fire behavior, our large-scale focus suggests that meteorological and fuel conditions are the primary drivers of intraday burning patterns. This also underscores the combined role of both fast-reacting meteorological conditions and slow-reacting fuel status in shaping hourly burning probability and fire persistence.

Trends in annual PBH

Using historical ERA5 fields and the final model trained on all 2017–2023 data, we reconstructed PBH within fire seasons for 1975–2024 across North America’s burnable areas at 0.25° (Materials and Methods). To assess extrapolation risk (49), we compared fire-season feature distributions between the historical (1975–2016) and training (2017–2023) periods for every combination of biome and variable (Materials and Methods). Within fire seasons, the historical feature space is almost entirely enveloped by the training period: median coverage of historical samples inside the training 0.1 to 99.9% band is 99.9% (98.4% for 1 to 99%; 100.0% for minimum-maximum; figs. S8 and S9, and table S3). This indicates our reconstruction is performed within the observed domain rather than by extrapolation into unseen conditions.

Overall, 52% of burnable area across North America showed significant increases in annual PBH (Mann-Kendall test; P < 0.05) during 1975–2024 and <0.5% exhibited decreases based on pixel-wise trend analysis. The total annual PBH rose by ~36% over 50 years, with western North America showing the most pronounced acceleration (Fig. 2). States such as California, Oregon, and Colorado saw widespread significant positive trends (76 to 93% of burnable areas), gaining an average of 11, 9.6, and 9.8 PBH/year, respectively (Theil-Sen slope estimator; table S4). In the Southwest (e.g., Arizona and New Mexico), ~90% of burnable areas added 13 to 14 PBH/year, with some grid cells reaching up to near 40 PBH/year (table S4).

Fig. 2. Significant trends in annual PBH, 1975–2024.

Map of significant PBH trends (Mann-Kendall, P < 0.05) per grid across burnable areas using Theil-Sen slopes (hours/year) based on RF modeling. White indicates no significant trends exist; light gray outlines states/provinces/territories. Inset: Total annual PBH per year with linear regression (red).

A latitudinal gradient emerged in both the magnitude and spatial extent of changes. Temperate and subtropical regions showed larger and more extensive gains than the boreal zone (table S5). For example, subtropical mountains and deserts gained an average of 16.9 and 13.4 PBH/year, respectively, across ~90% of areas (table S5); boreal areas in western Canada (e.g., British Columbia, Alberta, and Yukon) increased by 4 to 5 PBH/year across 66 to 89% of their burnable area (table S4). Notably, these trends may be moderated by regional variations in fuel availability and vegetation patterns (50).

Trends in seasonal PBH

Spring (March to May), summer (June to August), and fall (September to November) PBH increased consistently. Summer exhibits the largest absolute significant gains in seasonal PBH—tied to its larger baseline with 24% increases (30% of burnable areas; fig. S10). Spring and fall (46 and 34% of burnable areas) display steeper relative gains with 57 and 48% PBH increases during the past 50 years, respectively, suggesting that the shoulder fire seasons are more conducive to fires in recent decades in North America. Notably, our spring analysis shows that many of the significant increases cluster in western boreal areas (fig. S10). By contrast, in fall, there was a relatively higher proportion of significant positive trends in the eastern portion.

Rising fire risk from shifting diurnal pattern

We further quantified key metrics for each year during 1975–2024: the number of potential active days (days with ≥1 PBH), mean daily PBH, and two measures of extreme PBH days (≥12 or 24 PBH). The ≥12 PBH threshold represents an extension beyond normal diurnal rhythms, as the highest mean daily ABH observed across all biomes remains below 11 hours (table S1). The 24 PBH threshold characterizes extreme, continuous day-and-night burning activity. We found that rising annual PBH reflects both more potential active days and intensified diurnal burning potential. Across burnable areas, 44% exhibited significant increases in potential active days (mean cumulative gain +26.2 days over 1975–2024; ~+0.52 days year⁻¹; Fig. 3A); while 34% of areas showed a significant increase in mean daily PBH, averaging a 1.2 PBH gain for each potential active day during 1975–2024 (Fig. 3B). This pattern is particularly evident in historically fire-prone regions such as the subtropical mountain system, temperate mountain system west, boreal conifer west, and boreal tundra woodland west (table S6).

Fig. 3. Significant trends in potential active days and mean daily PBH, 1975–2024.

Map of significant trends (Mann-Kendall, P < 0.05) in annual potential active days (A) and mean daily PBH (B) per pixel across burnable areas, using Theil–Sen slopes (50-year totals) based on RF modeling. White indicates no significant trends exist; light gray outlines states/provinces/territories.

Specifically, in the temperate mountain system west and subtropical mountain system, mean daily PBH rose 22 and 21% during 1975–2024, respectively (Fig. 4, A and B), with the annual count of potential active days increasing by 29 and 34% concurrently. The surge in extreme PBH days was disproportionate. In boreal tundra woodland west, days with ≥12 PBH and 24 PBH increased by 81% and 233%, respectively (Fig. 4C). Similarly, in temperate mountain system west, these increases were 86% and 225%, respectively (Fig. 4A). These shifts underscore the dual risk of both longer fire seasons and more prolonged intraday burning.

Fig. 4. Shifting hourly pattern in fire-prone biomes, 1975–2024.

Changes in potential active days, mean daily PBH, and extreme PBH days (≥12 or 24 PBH) across significant areas in temperate mountain system west (A), subtropical mountain system (B), and boreal tundra woodland west (C) from 1975 to 2024 based on RF modeling. For each subplot, slope and delta indicate absolute changes per year and relative changes in total, respectively. Changes in other biomes are in table S6.

Parallel analysis

Last, to ensure our findings were not dependent on the choice of a specific machine learning algorithm, we conducted a comprehensive robustness check using a mechanistically distinct logistic regression (LR) model (Materials and Methods). This parallel analysis reconfirmed all our study’s central conclusions of PBH trend, including the direction and spatial pattern of the long-term increase in PBH, as well as the underlying shifts in seasonal and diurnal fire potential (figs. S11 and S12). The strong agreement between the two models provides evidence that the observed trends are a robust feature of the changing fire weather regime.

Field significance

A block-permutation field test for annual PBH during 1975–2024 (3-year blocks; 1000 permutations; Materials and Methods) confirms that the observed fraction of significant grid cells (52.14% for RF; 53.51% for LR) far exceeds the 99th percentile of the respective null distributions (31.65%/32.45%; right-tail field P < 0.001 for both models; fig. S13, A and B), establishing a field-significant, predominantly positive continental trend. Seasonal tests give similar conclusions (fig. S13, C to H).

DISCUSSION

Shifts in hourly fire pattern

While annual, seasonal, and daily fire activity are well-studied, diurnal burning patterns have long remained underexplored. However, recent widespread intraday surges across North America—e.g., continuous overnight burning and rapid high-intensity spread—have overwhelmed suppression efforts and devastated multiple communities, underscoring the urgency of this research. Spanning recent observations of ABH (2017–2023) and half-century trends in PBH (1975–2024), our analysis reveals a key mechanism driving North American fire regime transformations—climate-driven weakened constraints of the diurnal weather cycle lead to longer, more intense burning hours.

Our models indicate fast-reacting variables combination accounts for the largest share of performance gain, with their diurnal fluctuations traditionally creating natural constraints on fire activity (typically active day, quiet night). The weakening of these constraints is evident in our findings of increased daily PBH and more potential active days annually. As relative humidity and temperature form the complementary, first tier of drivers, this weakening likely manifests as reduced recovery in humidity and asymmetric diurnal warming (51, 52), allowing fires to maintain energy and spread potential during traditionally quiescent periods. It should be noted that this weakened diurnal constraint could arise from a rise in the mean baseline climate, a change in diurnal amplitude, or a combination of both. The analysis presented in this study captures the integrated outcome of these shifts and does not explicitly distinguish between these potential underlying drivers. This basic cycle shift cumulatively produces broader effects such as longer fire seasons and amplified peak burning periods. Addressing these challenges will require innovative approaches in fire science and management that account for the changing temporal dynamics of wildfires at hourly scales.

The substantial increase in annual PBH across North America over the past five decades reveals a continental-scale shift in fire-conducive conditions, with the western regions being particularly affected. This long-term modeled trend aligns with both documented increases in multiple fire regime characteristics (1, 3, 11, 13, 18, 19, 45, 53) and our recent satellite observations of a concentrated ABH in western mountains and boreal areas. Seasonally, shoulder seasons (spring and fall) have experienced proportionally greater PBH increases than summer, suggesting a substantial expansion of the traditional fire season. The spatial concentration of spring PBH increases in western boreal areas aligns with recent early-spring large fire observations in boreal regions (54, 55). Moreover, summer exhibits longer daily ABH across most biomes, likely driven by longer day length and frequent extreme fire-conducive conditions. Western boreal biomes exhibit their longest daily ABH in spring, which may be attributed to early snowmelt and rapid fuel drying before vegetation green-up, a “spring window,” when winds and continuous fuel availability override diurnal moderation (17, 54, 56). These seasonal and regional patterns highlight how climate change interacts with local ecological conditions to reshape fire regimes in distinct ways across North America.

Behind the annual and seasonal increases lie fundamental changes in diurnal burning potential. Many biomes, especially western mountains and boreal areas, have experienced substantial increases in both potential active days and mean daily PBH. The disproportionate surge in extreme PBH days marks a concerning shift in fire behavior, aligning with extreme fire weather (57) and the most intense fire events (45) surging in western North America and with our 2017–2023 observations of frequent continuous burning. This disruption in temporal patterns is further compounded by increasing intensity, as prolonged daily ABH is correlated with higher maximum intensities. Therefore, the weakening constraints in diurnal weather cycles not only extend burning into traditionally quiescent periods but also intensify fire behavior throughout active periods. Notably, ~60% of fires reached their maximum intensity within the first 24 hours after detection with ~17% being single-day fire events. This early peaking of fire intensity, combined with the early onset of overnight burning (17) and rapid spread (18), emphasizes the critical importance of rapid initial attacks and suggests that the window for successful containment may be narrowing. The convergence of early peaking intensity with extended burning hours also presents additional challenges.

A latitudinal gradient in PBH trends shows temperate and subtropical regions with larger, more extensive gains than boreal zones. Higher latitudes, with shorter fire seasons, may make modest absolute PBH increases proportionally more pronounced when viewed as percentage changes. Northern ecosystems may also require longer periods before climate warming signals fully manifest in fire weather conditions due to distinct seasonal patterns and temperature thresholds (58). As warming continues, this pattern may also foreshadow future changes in boreal regions toward conditions similar to the record-breaking Canada’s 2023 fire season, which maintained typical ignition numbers but increased ABH, resulting in burned area seven times the historical average (59). Notably, our analysis extending through 2024 reveals more extensive and pronounced trends in boreal areas than earlier studies covering periods ending in 2020 (57), suggesting that recent years have pronouncedly accelerated these trends. This also indicates a potential continued northward migration of extreme fire dynamics in coming decades, as evidenced by recent record-breaking polar and tundra fires (60, 61).

Our reconstruction targets weather-conditioned PBH—hours when local fire weather favors active burning. PBH does not encode changes in fuels, ignition pressure, or suppression capacity. Over the same period, however, factors such as fuel build-up from fire exclusion and limited prescribed burning, drought/insect legacies altering fuel structures, and rapid wildland-urban interface expansion and human activity can amplify or localize impacts. For example, human-caused fires have been shown to ignite under more extreme fire-weather and to be more severe than lightning-caused fires in California (62). We therefore interpret PBH trends as the weather pathway while acknowledging that fuels/people modulate where and how those hours translate into growth and damage.

Collectively, these findings indicate that a weakening constraint in the day-night weather cycle is reshaping both the duration and intensity of burning at operational timescales and is cumulatively driving broader fire regime transformation. Future management requires adaptation to wildfires that increasingly defy diurnal norms.

MATERIALS AND METHODS

Study area and biome categorization

Our study encompasses the continental United States and Canada using a comprehensive biome classification system adapted from (63) and building upon methodologies established in (57). This classification divides the study area into 16 distinct biomes, each characterized by consistent patterns of climate and vegetation. The classification includes three major boreal categories (coniferous forest—subdivided into west and east regions, tundra woodland—west and east, and mountain system), five temperate classifications [continental forest, oceanic forest, mountain systems (west and east), steppe, and desert], and five subtropical divisions (humid forest, dry forest, mountain system, steppe, and desert). The classification is completed with individual categories for tropical, polar, and water regions. Burnable areas in this study indicate all biomes combined except polar and water regions. Given that wildfire activity in the Temperate Mountain System predominantly occurs in its western portion, our analysis specifically focuses on this western North American region when discussing this biome.

Wildland fire geospatial databases

Our analysis integrates data from three wildland fire geospatial databases, each providing crucial information about spatial perimeters, burned areas, start dates, and end dates across North America. The first source, the Canadian National Burned Area Composite (NBAC) (64), forms an integral component of the Fire Monitoring, Accounting and Reporting System. This dataset, jointly developed by the Canada Centre for Mapping and Earth Observation and the Canadian Forest Service, provides burned area mapping from 1972 to 2023 (version 20240530). NBAC achieves high spatial resolution through a combination of 30-m Landsat imagery and even higher-resolution agency imagery (<30-m spatial resolution).

For US fire data, we primarily relied on the Monitoring Trends in Burn Severity (MTBS) (65) database, supplemented by the Interagency Wildland Fire Perimeter History (IWFPH). The MTBS, administered jointly by the US Geological Survey and the USDA Forest Service, documents burn severity and extent across the United States from 1984 to 2022 (released 22 August 2024). This database uses different size thresholds for inclusion: fires ≥1000 acres (~405 ha) in the western United States and ≥500 acres (~202 ha) in the eastern United States. To ensure comprehensive coverage through 2023, we complemented the MTBS data with fire perimeter information from the IWFPH (version 20240825), maintained by the Wildland Fire Management Research, Development, & Application program data team and provided by National Interagency Fire Center.

Geostationary active fire detections

Consistent with (17), we used the Fire/Hot Spot Characterization Full Disk (FDCF) products from the GOES-R series (i.e., GOES-16, GOES-17, and GOES-18) to derive subhourly active fire detections across North America during 2017–2023. GOES-16 (GOES-East) has been stationed at 75.2°W over the equator since 2017; GOES-17 (GOES-West) operated near 137.2°W from late 2018 until early 2023; GOES-18 has occupied ~137.2°W since late 2022 and replaced GOES-17 as the West satellite in 2023. Therefore, the periods of these three FDCF products used in our study were: GOES-16, first available date to 2023; GOES-17, first available date to 2022; and GOES-18, 2023. Used together, GOES-East and GOES-West provide geostationary coverage of all burnable lands in North and South America. FDCF uses visible and infrared Advanced Baseline Imager (ABI) bands to locate fires and retrieve subpixel fire properties at 5- to 15-min temporal resolution and ~2-km nominal pixel size at the subsatellite point [effective resolution coarsens with view-zenith angle (VZA)] (29).

Notably, the availability, frequency, and quality of the GOES-R FDCF varied both regionally and over the study period. This is a result of the sequential launch of GOES-16, GOES-17, and GOES-18; changes in imaging frequency; the specific coverage area and VZA of each instrument; and short, sporadic instrument outages. Accordingly, we restrict our use of FDCF datasets here to identify the binary hourly burning status (active or inactive), rather than hot spot counts, for fires with well-defined spatio-temporal boundaries from wildland fire databases. As the available data provide at least four observations per hour at any location across the study area, we considered these data fit for this purpose.

Specifically, because of the sequential launch of these three satellites and the different coverage of GOES-East and GOES-West, the data availability is not consistent across study area during 2017–2023 (fig. S14). Northwestern North America (including Alaska and Yukon) was not visible to GOES-East, so its active fire detection data are only available from August 2018 onward, when GOES-17 (and later GOES-18) became operational. Conversely, the northeast was only imaged by GOES-East. In the central region where coverage overlapped, the data volume approximately doubled after GOES-17/18 was active.

Moreover, the scanning mode of each satellite has also changed over time during 2017–2023 to meet changing operational and experimental needs. This led to varying frequency of full disk imagery and thus FDCF products derived from it between 5 min (mode 4), 10 min (mode 6), and 15 min (mode 3). As a result, in the early years of the mission (2017–2018), GOES-16 and GOES-17 operated mostly in mode 3 (15-min mode). From 2019 onward, mode 6 (10-min mode) became the dominant operational mode for these two satellites and for GOES-18 when operated. Therefore, a higher volume of FDCF data is available in the later years of the study period (fig. S14). Because ABH is derived at an hourly resolution as a binary indicator of whether any active fire was detected within the fire perimeter during a given hour, the pre-2019 15-min sampling still provides multiple observations per hour and should be sufficient to classify hourly activity. The increased sampling after 2019 primarily adds redundancy within the same hour rather than systematically increasing the number of ABH; however, it is notable that this could slightly increase the probability of detecting very short-lived or near-threshold fire signals. For more information on GOES-R scanning mode, please refer to www.goes-r.gov/.

Active fire detections from GOES-16, GOES-17, and GOES-18 were therefore pooled to create a single, continental-scale dataset. Using them together ensures the most complete geostationary observational record possible for any given fire location across North America. Our goal was to maximize detections for each fire event, not to intercompare the performance of the individual satellites.

There are some technical challenges of using GOES data. First, because GOES-R satellites are in geostationary orbits, the VZA for any ground point is fixed relative to each instrument. VZA affects the accuracy of fire detection algorithms in several ways (66, 67). For our study, the fact that the ABI pixel size (i.e., spatial resolution) and hot spot omission error rate increase with VZA—especially for small/cool fires and at high latitudes and complex terrain. This consideration leads us to analyze binary burning status rather than hot spot counts and avoid performing detailed direct diurnal fire cycle intercomparisons between biomes with very different VZAs. Geostationary viewing geometry (more pronounced at larger VZA) and terrain effects can also cause parallax error with displacement of pixels. To mitigate geolocation and fire perimeter uncertainties, we used a 2-km buffer (approximately nominal ABI fire-pixel width) around each fire perimeter when associating hot spots with a specific fire (see the “ABH and intensity metrics” section). Moreover, relative to daytime, nighttime backgrounds are typically cooler and more homogeneous, generally increasing the potential contrast provided by active fire pixels. The nighttime active fire detection algorithm is therefore considered to be more sensitive, especially to smaller and/or cooler fires than the daytime algorithm (68).

ABH and intensity metrics

We developed a stepwise methodology to analyze hourly burning status (active or not) and intensity across multiple temporal scales. The process began with the extraction of GOES-R active fire hot spots that intersected with fire perimeters from our combined database (NBAC and MTBS-IWFPH). To associate hot spots with events, we intersected GOES-R hot spots with perimeters buffered by 2 km (~nominal FDCF pixel width), which accommodates ABI geolocation error, parallax at oblique angles, and perimeter uncertainty. A hot spot intersecting with the buffered perimeter was considered event-associated. To establish the temporal boundaries of fire events, we used a dual approach: utilizing recorded start and end dates when available or inferring these dates from GOES-R hot spot data based on the first and last instances of consecutive ABH.

To ensure temporal accuracy, we converted all fire activity data from universal time coordinated to local time zones using the spatial centroid of fire perimeters and accounting for day of year. We implemented precise daylight categorization by using exact sunrise and sunset times, defining sunrise as the moment when the sun’s top edge appears on the horizon and sunset as its disappearance below the horizon. This allowed us to categorize each hour into one of four states: active daytime, nonactive daytime, active nighttime, or nonactive nighttime. For analytical purposes, we defined a day as the period from the first hour after sunrise to the last hour before the next sunrise and defined each hour as a half-open interval [hh:00:00, hh:59:59)—daytime spans from the first hour after sunrise to the last hour before sunset, and hours outside this range are nighttime. The event temporal window is therefore from the first hour of the start date to the last hour of the end date.

For each fire event, as abovementioned, we extracted all hot spots that occurred with the 2-km buffered perimeter between the event’s start and end dates and used these to determine a binary hourly status. Theoretically, within this spatio-temporal window, the physical fire state is either flaming or smoldering most of the time; however, when status needs to be determined from the geostationary perspective, “nonactive” hour (no hot spot reported) may reflect nonflaming/smoldering combustion below the detection threshold, very small flaming area, cloud/canopy obstruction, unfavorable viewing geometry, or genuine inactivity (68). Except for the data quality flag (i.e., only keep “good quality fire”) from FDCF, we therefore did not apply additional quality mask beyond the spatio-temporal window. We used GOES-R FDCF level 2 files distributed via the National Oceanic and Atmospheric Administration (NOAA)’s AWS Open Data repositories (https://registry.opendata.aws/noaa-goes/). Short, sporadic satellite anomalies and outages do occur (for more information, please see www.ospo.noaa.gov/operations/goes/status.html); however, our analysis aggregates over hundreds of thousands of fire hours (with modeling and most analysis carried out at the fire-day level), so these gaps are unlikely to introduce systematic bias in the analysis and trends.

We excluded fires from 2017 to 2023 that lacked corresponding hot spot data. We established specific criteria for fire activity and intensity measurement. A fire was classified as active at any hour where at least one hot spot was detected within its perimeter. We then summarized the mean ABH for each combination of biome and season.

To quantify burning intensity, we used the FRP values accompanying each GOES-R hot spot. We focus on two metrics that capture fire intensity at different time scales for each fire within its perimeter: daily and fire-event maximum FRP. Specifically, the daily and fire-event maximum FRP are the highest among all hourly maximum FRPs recorded in 1 day and throughout the entire fire event, respectively. Maximum FRP can serve as an indicator of extreme burning intensity and particularly corresponds to head fire behavior—a critical driver of explosive and destructive spread (46, 69). High FRP often correlates with fast-moving flame fronts, increased convective energy, and high heat flux, all of which challenge suppression efforts and increase risks to firefighting resources. We then analyzed the relationship between daily ABH and maximum daily FRP values. At the event level, we examined the correlation between cumulative ABH and maximum FRP using Spearman’s rank correlation coefficient. We also investigated relationships between fire size, cumulative ABH, and fire event maximum FRP. For these between-fire comparisons, we conducted a sensitivity analysis using a pixel footprint area–normalized fire-event maximum FRP metric (MW km⁻²). Pixel footprint area was estimated from viewing geometry using detections’ latitude/longitude and GOES-R platform (GOES-16/17/18), and we used the fire-event maximum of the resulting FRP density for the between fire analyses. In addition, we measured the time between the fire-event maximum FRP and the first detection date for each fire event and analyzed the frequency of nighttime peak intensities for each active day if burning into the night. All temporal calculations were facilitated through the R packages lutz and suncalc, ensuring precise time zone conversions and solar position determinations.

We used the Earth observation fire products from the GOES-R FDCF because it is the only platform providing consistent, high-frequency (≤15-min) fire detections for North America, making it uniquely capable of capturing hourly fire patterns. However, the reported number and characteristics of burning hours are likely to be underestimated, given the omission errors inherent in active fire detection algorithms (68). Moreover, we use FRP as a relative intensity indicator. Radiometric saturation at very high subpixel temperatures can underestimate extreme FRP, so daily or event-maximum FRP may be conservative—particularly on days with long ABH.

Fire weather calculation and extraction

Fire weather variables were derived on the basis of ERA5 reanalysis and the CFWIS, a widely used operational framework for assessing fire danger that typically calculates six components, including three fuel moisture codes [Fine Fuel Moisture Code (FFMC), Duff Moisture Code (DMC), and Drought Code (DC)] and three fire behavior indices [Initial Spread Index (ISI), Buildup Index (BUI), and Fire Weather Index] (48). ERA5 provides global atmospheric fields on a 31-km horizontal grid with hourly resolution, enabling us to compute high-frequency meteorological variables.

From ERA5, we extracted or calculated four fundamental meteorological variables that serve as inputs of the CFWIS: hourly precipitation, wind speed, 2-m temperature, and relative humidity. Following established procedures detailed in (17), we used these variables to calculate six CFWIS components that characterize different aspects of fire potential. For our analysis, we maintained certain variables at hourly resolution (temperature, relative humidity, wind speed, precipitation, FFMC, and ISI), while others were computed at daily steps (DMC and BUI) to capture both rapid and cumulative changes in fire environment conditions.

For each fire event, the extracted variables were spatially averaged across all ERA5 grid cells that intersected the fire perimeter. The extraction period spanned the entire lifetime of each fire, with a 24-hour buffer added at the start to capture early fire conditions. Last, these temporally aggregated fire weather metrics were time-matched to the corresponding hourly fire dynamics used in our analyses.

Modeling hourly fire dynamics

We developed a machine learning framework to estimate the hourly burning status of fire events during active days. Our model, implemented as a RF classifier, incorporated both hourly and daily fire weather variables derived from our ERA5-CFWIS methods. To capture environmental and seasonal variations, we included additional categorical variables—biome classification and month—which were integrated into the model through one-hot encoding. We used RF because it is an ensemble model, which is built on multiple decision tree classifiers and can capture complex, nonlinear relationships between fire weather and burning status with robust performance.

The model configurations are as follows: We implemented a stratified threefold cross-validation strategy, which maintained consistent class distributions across training and validation sets. To address the inherent class imbalance in fire activity data, we applied a balanced weighting scheme during model training. Each RF was constructed with 500 trees, with a maximum of three features considered at each split decision to prevent overfitting while maintaining model complexity sufficient for capturing underlying patterns in fire behavior.

We evaluated the model performance under three training/test splitting approaches: (i) fire-hour split: Each hour was split into either training or test set; (ii) fire-day split, where all hours for a fire on a sunrise-to-sunrise day are withheld together; and (iii) fire-event split, where all hourly data from a given fire event are entirely withheld in either train or test set. Model optimization involved determining the optimal classification threshold through precision-recall curve analysis, specifically maximizing the F1 score to balance precision and recall. We evaluated the model’s binary classification performance using five standard metrics. The area under the receiver operating characteristic curve (AUC) was calculated to measure the model’s overall ability to distinguish between active and non-active hours. Sensitivity (recall) quantified the proportion of true ABH that were correctly identified by the model, while specificity quantified the proportion of true nonactive hours that were correctly identified. Accuracy indicates the proportion of all predictions that were correct. Last, the F1 score was calculated as the harmonic mean of precision and sensitivity.

Model interpretation and multicollinearity

Because several variables are correlated, we avoided overinterpreting single-variable ranks and instead used diagnostics that are more stable under multicollinearity: (i) group-wise permutation, (ii) group ablation with retraining, (iii) TreeSHAP (70), and (iv) accumulated local effects (ALEs) (71) for RF model with fire-hour split. Groups were: Hourly_raw = (temp, rh, winds, prec, ffmc, isi), Temp + RH = (temp, rh), Hourly_index = (ffmc, isi); and Daily_drought = (BUI, DMC). All diagnostics used the same cross-validation splits as model evaluation. For permutation, features within a target group were permuted jointly in the held-out fold (eight permutations per fold); ΔF1 and ΔAUC were computed relative to the unperturbed model. For ablation, the group was removed from train/validation, the model retrained on the same splits, and ΔF1/ΔAUC reported. TreeSHAP values were computed on held-out predictions with path-dependent semantics and summarized as mean |SHAP| for the positive class (“ABH”). ALE curves (Temp, RH) were computed on held-out data over the empirical range and centered (ALE = 0 at the overall mean prediction).

Long-term trends in hourly fire potential

To analyze historical trends in hourly fire potential across North America, we applied our calibrated RF classifier (fire-hour split, trained on all 2017–2023 data) to fire weather variables from ERA5-CFWIS methods spanning 1975–2024 at a spatial resolution of 0.25°. Our analysis is strictly confined to fire seasons, which are defined annually for each grid cell based on physical temperature thresholds (fire season starts when three consecutive days have a maximum temperature above 12°, and fire season ends when three consecutive days have a maximum temperature below 5°). We defined “PBH” as hours where the model-estimated probability of active burning exceeded our optimally determined threshold, providing a consistent metric for comparing fire potential across space and time.

To evaluate long-term trends in fire activity, hourly estimations were aggregated into annual and seasonal total PBH. This pixel-wise analysis employed a multiscale approach, examining patterns at continental, political (US states and Canadian provinces/territories), and biome levels. We used the Mann-Kendall test (P < 0.05) to assess the statistical significance of trends, while the magnitude of changes in annual PBH was quantified using Theil-Sen regression, chosen for its robustness to outliers and non-normal distributions.

To investigate changes in diurnal fire patterns, we further quantified three key metrics for each year during 1975–2024: the number of potential active days (with at least one PBH), mean daily PBH, and the frequency of extreme PBH days (≥12 or 24 PBH). Similarly, Mann-Kendall test and Theil-Sen regression were used for each metric in each grid and then summarized for each burnable biome. This comprehensive analytical framework allowed us to differentiate between changes in fire season length and shifts in diurnal fire potential, providing insight into the evolving nature of fire risk across North America.

Distribution overlap and extrapolation check

We assessed whether the historical feature domain (1975–2016) is contained within the training domain (2017–2023) within the fire season (season definition given above). Variables considered were Temp, RH, WS, Prec, FFMC, ISI, BUI, and DMC, evaluated separately by biome. For each biome × variable, we randomly sampled up to 100,000 fire-season observations from the historical period and the same number from the training period. Our primary metric is the coverage of historical samples inside the training 0.1 to 99.9% range (“coverage” hereafter). As sensitivity checks, we also computed coverage against the training 1 to 99% range and against the training minimum-maximum range.

Parallel LR analysis

We trained a regularized logistic regression (LR) model as a mechanistically distinct baseline to the RF model. LR used the identical predictor set, preprocessing, and cross-validation splits as RF with feature standardization. We then searched 50 randomly sampled settings (L1 or L2 penalty, regularization strength, and solver) and chose the setting with the best F1 score. Out-of-fold performance with fire-hour split was: AUC = 0.76, sensitivity = 0.75, specificity = 0.63, accuracy = 0.67, and F1 = 0.58. For PBH reconstruction, we applied the full trained LR model to 1975–2024 and aggregated hourly predictions in the same way as for RF. Trends were assessed per grid cell with the Mann-Kendall test (α = 0.05); slopes were estimated with Theil-Sen. As a result, the high degree of consistency in trend direction, spatial patterns, and temporal dynamics (figs. S11 and S12) between RF and LR (with RF showing stronger performance) provides strong evidence that our central findings are robust and reflect a fundamental signal in the climate data, rather than being an artifact of a specific model choice. Therefore, we report the results from the primary RF analysis, while this LR analysis serves as a confirmatory robustness check.

Field significance test

To control for multiple testing across grid cells, we assessed field significance using block permutations of the temporal order. Annual and seasonal PBH time series for both RF and LR models at each grid cell were kept intact within 3-year blocks (chosen to preserve short-term persistence), and blocks were randomly permuted 1000 times to form null realizations with no coherent trend. For each permutation we ran per-pixel Mann-Kendall tests (α = 0.05) with trend sign from Theil-Sen and computed the fraction of significant pixels within the study area. The field P value was the right-tail probability of the observed fraction under the permutation distribution.

Supplementary Materials

This PDF file includes:

Figs. S1 to S14

Tables S1 to S6