1. Introduction
Wildfires have become increasingly common and destructive in many regions of the world, posing a significant threat to human life, critical infrastructure, and natural resources (Jolly et al. 2015; Lelieveld et al. 2015; Higuera et al. 2023). To manage the risk, wildfire agencies rely on fire danger rating systems (FDRSs) to assess the potential severity of wildfires in a particular area and to assign a numeric or categorical rating (Andrews et al. 2003). FDRSs consider various factors such as weather conditions, fuels, fuel moisture levels, topography, and historical fire behavior. This rating is typically given for a particular location to estimate the potential for a wildfire to start, spread, and become difficult to control (Canadian Forest Service Fire Danger Group 2021).
FDRSs play a significant role in determining the appropriate level of fire suppression resources needed, including personnel and equipment. FDRSs can also be used to issue fire restrictions or bans for certain regions, which can help reduce human-caused fires. Additionally, FDRS forecasts are crucial for allocating these fire suppression resources strategically across a country, ensuring that personnel and equipment are positioned where the risk of wildfires is highest. This proactive approach enhances the efficiency and effectiveness of firefighting efforts. FDRSs also provide valuable information to the public, including fire danger warnings and advice on fire protection measures, contributing to overall wildfire management and prevention.
One component of the Canadian Forest Fire Danger Rating System (CFFDRS) is the fire weather index (FWI), which assesses the fire danger level based on weather conditions (Van Wagner and Pickett 1985; Van Wagner 1987; Forestry Canada 1992; Lawson and Armitage 2008). Even though it was originally developed empirically to estimate fire behavior in a typical Canadian forest ecosystem, predominantly consisting of red and jack pine stands, FWI has been shown to be an effective rating system globally (Van Wagner 1974, 1987; Taylor and Alexander 2006; Cruz and Plucinski 2007; Groot et al. 2007; Bedia et al. 2012; De Groot et al. 2015; Field et al. 2015; Di Giuseppe et al. 2016). FWI is also used in some U.S. states, including Alaska and several Great Lakes states. It has also been successfully used as an aid to improve estimations of wildfire smoke emissions by utilizing fire radiative power (FRP) observations (Di Giuseppe et al. 2017, 2018).
Components of the FWI system include the fine fuel moisture code (FFMC), duff moisture code (DMC), and drought code (DC), which describe the moisture content of organic material on the forest floor at various depths. Each code has a unique response time with respect to current weather conditions; the FFMC responds the most quickly, followed by the DMC and then the DC, as indicated by their respective equilibrium drying rates of 2/3, 15, and 53 days (Lawson and Armitage 2008). The moisture codes function as a tracking system to monitor how local weather conditions affect fuel moisture over time by iteratively applying the previous day’s values as input to the current day’s calculations. Each code has a specific default start value when data from the previous day are not available.
Other outputs of the FWI system include the initial spread index (ISI), build-up index (BUI), and the FWI. The ISI quantifies the rate at which fires could propagate during their initial stages. The wind speed combined with the weather-dependent FFMC is used to define this index. The BUI quantifies the amount of fuel available for combustion and is determined by DMC and DC values. The FWI then amalgamates the ISI and BUI values to estimate the general fire danger (Fig. 1).
The structure of the Canadian FWI system.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
For more detailed information on how to calculate, interpret, and gain a deeper understanding of the FWI system, please refer to Van Wagner (1987) and Lawson and Armitage (2008).
All components of the traditional FWI system characterize conditions at an assumed midafternoon period of peak fire activity, approximately 1600 local time. These conditions are derived from four input weather variables: 2-m temperature, 2-m relative humidity (RH), 10-m wind speed, and 24-h accumulated precipitation, recorded at standard noon local time. The choice of standard noon local time is based on historical weather stations that recorded data at this specific time (Van Wagner 1987; Lawson and Armitage 2008).
While the fundamental concepts of the FWI system are widely applied globally, FWI forecast implementation techniques differ between countries. The current operational FWI system in Canada is run by Natural Resources Canada [NRCan; Natural Resources Canada Canadian Wildland Fire Information System (2023)]. Their method spatially interpolates the four input weather variables, valid at standard noon local time, from North American weather station locations to a 1 km × 1 km gridded domain covering all of Canada, using inverse distance weighting. Next, temperature and relative humidity are elevation-adjusted using a standard lapse rate of 6.5°C km−1. The six indices/codes that comprise the system are then calculated for every grid cell, creating maps of the FWI output.
For “day 0” of the FWI forecast, the FWI values are calculated directly from the station observations. On subsequent forecast days (“day 1” onward), point location weather forecasts at the station locations are interpolated to the same 1 km × 1 km grid. These forecast variable fields are input to the FWI equations along with the moisture codes from day 0, to help keep the FWI forecast close to observations within the iterative process. This is repeated throughout the forecast, with the previous day’s observation-based moisture codes fed into the next day’s FWI forecast along with that same day’s numerical weather prediction (NWP) point location forecasts interpolated to the grid. For forecast days 1 and 2, the Global Environmental Multiscale (GEM) model is used, and for days 3–14, the North American Ensemble Forecast System (NAEFS) is used—offering 14-day, daily FWI forecast maps for Canada.
However, due to the large spatial extent of Canada (10 million km2) and the paucity of weather stations especially further north, this approach leaves large and nonuniform distances, which contain diverse terrain between the interpolated weather locations. As a result, critical fire weather data between locations are often missed or incorrectly applied, particularly in complex terrain. To mitigate this issue, newer FWI forecast products have been developed by directly using NWP models to derive and iteratively track the FWI system at every surface model grid cell. Several groups have shown the skillfulness of this approach, especially in areas with sparse observation locations (Mölders 2010; Di Giuseppe et al. 2016; Vitolo et al. 2020; Mandal et al. 2022).
All the previously discussed methods, however, only use standard noon local weather to derive a daily fire danger rating representing peak fire activity assumed to occur at a midafternoon time of approximately 1600 local standard time (Van Wagner 1987; Lawson and Armitage 2008). However, peak fire activity does not always occur at this time. Also, as human-caused climate change alters forest fuels and the diurnal wildfire cycle, intense fire behavior in the morning and late evening is expected to increase (Flannigan et al. 2016; Balch et al. 2022; Freeborn et al. 2022). This motivates the development of a subdaily FDRS to capture these changes and risks on this time scale.
Our study investigates the potential of an FWI forecast system with a 1-h temporal resolution using hourly output from the Weather Research and Forecasting [WRF; Skamarock et al. (2020)] NWP Model. It accomplishes this by calculating FFMC/ISI at hourly intervals using modified empirical formulas (Van Wagner 1977; Beck and Armitage 2004) compared to the conventional daily approach (Van Wagner 1987; Lawson and Armitage 2008) previously described. These hourly FFMC/ISI values are then combined with the slower-evolving components (DMC/DC/BUI) to generate an hourly FWI (HFWI) forecast.
This paper presents a comparison of the HFWI and daily FWI (DFWI) methods, both of which utilize the same WRF model output for this study. Each method is validated by comparing them with observation-based daily FWI values from 917 surface fire weather stations (see section 2 for definition) in western North America. To further examine performance, we compare each method with observations of FRP from geostationary satellites for nine wildfire case studies spanning Canada and the United States. Section 2 describes the methods, models, and data; section 3 presents the results; and conclusions are provided in section 4.
2. Methods
a. Weather inputs
Our NWP-derived hourly FWI system, hereafter referred to as the fire weather forecast (FWF) model, utilizes weather forecast data generated by the WRF Model [version 4.2.1; Skamarock et al. (2020)]. WRF is configured with three nested domains which have horizontal grid spacings of 36, 12, and 4 km (Fig. 2). The domains utilize two-way nesting, facilitating the bidirectional exchange of meteorological variables between consecutive spatial resolutions, providing a comprehensive representation of atmospheric processes at varying scales.
Map showing the spatial extent for the 36- (blue), 12- (red), and 4-km (green) domains. Also shown is the location of the 917 surface weather stations (gray dots) used to validate model performance.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
We run the described WRF configuration on the Google Cloud Platform (Chui et al. 2019) once daily with a total forecast horizon of 60 h, utilizing initial and lateral boundary conditions from the North American Mesoscale Forecast Model (National Centers for Environmental Prediction, National Weather Service, NOAA, U.S. Department of Commerce 2015), specifically for 0000 UTC with 32-km grid spacing. To account for model spinup, the initialization hour and first five forecast hours are rejected for each run, resulting in our FWI methods utilizing a forecast horizon of 55 h (Skamarock et al. 2020). This WRF configuration has been operational since 1 January 2021 and further information is provided online (http://weather.eos.ubc.ca/wxfcst/html-etc/model-metadata/wan00cg-01.html).
b. FWF model
The FWF model is written in Python and generates daily and hourly FWI forecasts. The DFWI is calculated at each grid cell location within the 12- and 4-km WRF domains using data valid at standard noon local time, accounting for time zone offset from UTC (Fig. 3). The time zone mask efficiently extracts standard noon local values for each surface grid cell through vectorized indexing. This method treats the unique time zone offsets from UTC as integers to concurrently index the gridded domains for each forecast day. The FWF model is configured to work for any model initialization time and will automatically adjust the indexing procedures accordingly for obtaining standard noon local weather. The FWF model can work with any modern NWP model output after running the setup script to create the required time zone offset from the UTC mask.
Map showing the UTC offset mask (during standard time) for the 12-km WRF domain.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
Next, to compute the HFWI values, the FWF model iteratively calculates the FFMC at an hourly interval using hourly WRF model output of air temperature, relative humidity, wind speed, and accumulated precipitation for each forecast day. The hourly ISI is determined by utilizing the hourly FFMC values along with hourly wind data from WRF (Van Wagner 1977). The FWF model then partitions the codes and indices of the FWI system that are less sensitive to weather variations, namely, the DMC, DC, and BUI, based on forecast day (day 1, day 2, etc.). This partitioning process utilizes the UTC offset mask to establish a 24-h, midnight-to-midnight local time range for each grid cell. Following the partitioning step, the FWF model employs the partitioned BUI and hourly ISI values to solve for an HFWI, for that same time period. Finally, utilizing the midnight-to-midnight local time information, the FWF model also determines the maximum FWI value in that time period and the local timing of its occurrence at each model grid cell.
The FWF model generates two Network Common Data Form (netCDF) output files containing the daily and hourly FWI system values along with all associated meteorological input parameters. Additionally, the saved fuel moisture code values are subsequently used as initial inputs to the next forecast day.
The FWF model was initialized on 1 January 2020 using default start-up values of 85, 6, and 15 for the FFMC, DMC, and DC, respectively (Van Wagner and Pickett 1985; Lawson and Armitage 2008). Next, 1 year of reanalysis data (1 January–31 December 2020) from the fifth major global reanalysis produced by European Centre for Medium-Range Weather Forecasts (ECMWF) (ERA5) model was used to spin up the fuel moisture codes (Hersbach 2023) from the initial values. The spun-up moisture codes were reprojected to the WRF Model domains and used as inputs on 1 January 2021.
Importantly, calculations of the moisture codes were not suspended during the winter months, allowing the FWF model to iteratively track the moisture codes throughout the year. This approach avoided overwintering of the moisture codes and the need for a user-defined stop–start date for the wildfire season (Lawson and Armitage 2008; Vitolo et al. 2020), instead relying on WRF to model the physical, dynamic weather processes and interactions that occurred during winter. This was important since human-caused climate change might reduce the duration of nonfire season (Jolly et al. 2015), resulting in different stop–start dates with different conditions.
The FWF model Python code, along with documentation detailing the configuration, execution, and working with the model data, is publicly available on GitHub under an open-source MIT license. The GitHub repository and documentation can be accessed online (https://github.com/cerodell/fwf/tree/ams-fwi and https://cerodell.github.io/fwf-docs).
All FWF codes and indices, associated meteorology, and current wildfire locations are displayed on a zoomable map with pop-up point forecast functionality (Fig. 4), updated daily (https://firesmoke.ca/forecasts/fireweather/).
An example from the FWF web page. The HFWI is shown valid at 1700 PDT 2 Jun 2023 (0000 UTC 3 Jun 2023). Also shown are the various observation and forecast products offered.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
c. Observations
1) Fire weather station data analysis
We obtained surface weather station observation data from the NRCan Canadian Wildland Fire Information System (CWFIS) to validate and compare the DFWI and HFWI forecast methods (Natural Resources Canada, Canadian Forest Service 2023). Our study period for this part of our analysis is two full Canadian wildfire seasons of 2021 and 2022 (currently defined as 1 April–31 October). These data include observations of 2-m temperature, 2-m relative humidity, 10-m wind speed, and 24-h accumulated precipitation (Fig. 1) from 917 surface weather stations, recorded once per day at standard noon local. It also includes observation-derived daily FWI values calculated using the CFFDRS R package (Wang et al. 2017) using the daily weather observations as input. These observation-derived DFWI values are used operationally by provisional wildfire agencies across Canada. Fire weather stations within the 12- and 4-km domains were selected for the analysis (Fig. 2), and only stations with continuous observations of more than 30 days were considered. This dataset is used for comparison against the WRF-derived HFWI and DFWI (described at the start of section 2).
2) Fire radiative power: Wildfire case study analysis
(i) Fire radiative power data
To further validate our FWF model, we compare it against satellite observations of FRP, a measure of the radiative energy emitted by an active burning wildfire, which is correlated to the total biomass combusted (Wooster et al. 2005). A comparison of FRP with FWI is meaningful, as FWI is meant to represent fire intensity, which is considered to be the amount of heat released from an active fire (Van Wagner 1987; Di Giuseppe et al. 2017). Di Giuseppe et al. (2017) identified a linear relationship between daily FWI, daily FRP, and wildfire emissions, further supporting our approach.
The FRP observations were from the fire detection and characterization (FDC) product derived from the Advanced Baseline Imager (ABI) on the R series of Geostationary Operational Environmental Satellite (GOES-R) from the National Oceanic and Atmospheric Administration (NOAA) (Schmidt et al. 2012). The GOES-R satellites GOES-East (GOES-16) and GOES-West (GOES-17 or GOES-18, date-dependent) are in fixed geostationary orbits of 75.2° and 137.2°W, respectively, and provide 15-min observations of FRP, enabling the diurnal variations in fire activity to be captured at cloud-free times over the study area (Schmidt et al. 2012). The FRP FDC product is relatively new and has been validated against higher spatial resolution polar-orbiting satellites yielding reasonable results, which provides confidence in our approach (Li et al. 2020).
To compare FRP with WRF-derived FWI values, first, we identified wildfire case studies within the 12- and 4-km model domains from 2021 to 2023 (Fig. 5) with the National Aeronautics and Space Administration (NASA) Worldview tool. Using visible true color images and thermal anomalies from the Visible Infrared Imaging Radiometer Suite (VIIRS) instrument aboard the joint NASA/NOAA Suomi National Polar-Orbiting Partnership (Suomi NPP) and NOAA-20 satellites, we were able to determine the approximate start date and time and geographical extent of the wildfires. An interactive NASA Worldview example of the case study of the Caldor Fire that burned near South Lake Tahoe, California, United States, in late August 2021 can be found online (https://go.nasa.gov/47K1bKh).
Map of wildfire case study locations (red triangles) used in this study and the three nested WRF domains.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
For each wildfire case study, FRP data were obtained from one of the two GOES-R satellites, determined based on which had the best incidence angle of observation, i.e., which satellite was more directly overhead. The observed FRP for a specific case was spatially averaged over the wildfire area estimated using thermal anomaly extent, considering only the processed fire pixels. These are pixels that have undergone rigorous postprocessing and are identified with the highest confidence level by the FDC algorithm (Schmidt et al. 2012).
Next, we time-averaged the 15-min frequency FRP fields to a 60-min frequency for direct comparison with HFWI output. The time period for each case study extended to a maximum of 10 days past the first thermal detection. Note that not all wildfire cases were burning or had continuous data available for the full 10 days, and cases with 12 or more hours of continuously missing data were eliminated, leaving nine cases in total (see Table 2 in section 3 for the full list of case studies).
As an example, Fig. 6 displays FRP data for the Caldor Fire (near South Lake Tahoe on 16–25 August 2021) case study. Figure 6a shows the spatial extent and timing of FRP observations, and Fig. 6b shows a time series of 15-min raw data along with the hourly averaged data used for comparison [i.e., hourly FRP (HFRP) time series data used in the comparison].
(a) Spatial field of FRP observations (MW) for the Caldor Fire, colored by the timing of the observation (15-min frequency) and (b) time series of 15-min spatially averaged raw data [MW; circles using the same color scheme as in (a)] and the hourly averaged data (black line) used for comparison with the HFWI from the FWF.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
(ii) Hourly fire weather index data
We compared our HFWI forecast with the HFRP time series (e.g., Fig. 6) for each wildfire case study by generating a single HFWI time series using the FWI values from the nearest grid cells to each FRP observation grid center location during the analysis period (e.g., Fig. 7). The nearest grid cells were then weighted based on the number of FRP detections within each FWI grid. Finally, the weighted FWI grid cell values were spatially averaged, producing a time series of HFWI data. The first 24 h of each HFWI daily spun-up forecast were concatenated to cover the full multiday comparison using the finer-resolution WRF domain where available. We also created an hourly time series of DFWI for visual comparison with the HFWI and HFRP, where the single noon local daily FWI forecast value is repeated for the corresponding midnight-to-midnight 24-h period.
Weighted WRF-derived FWI grid cell values for the Caldor Fire, valid from 1000 PDT 16 Aug to 1200 PDT 25 Aug 2021. WRF 12-km grid cells are represented by green and red boxes. Black dots overlaid on the WRF grid denote the FRP observational data. The WRF grid cells are colored based on the number of FRP observations within each specific grid. These color-coded weights are used for spatial averaging, providing the weighted time series of HFWI forecast data.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
d. Verification methods
To compare verification statistics for the HFWI and DFWI methods, assessing their performance against observation-derived HFWI for the two fire seasons, we used Taylor diagrams, which present Pearson’s correlation, root-mean-square difference (RMSD), mean bias error (MBE), and standard deviations (std dev) of the observed and modeled data.
Additionally, Pearson’s correlation coefficient was used to compare the HFWI prediction versus the observed HFRP data from the GOES-R satellites for each wildfire case study. The DFWI was resampled to hourly for an intercomparison of the two methods, as described previously.
3. Results and discussion
a. Fire weather station analysis
We include some important reminders and associated information at this point:
-
We compared the WRF-derived DFWI and HFWI forecast methods with the observation-derived DFWI values at 917 fire weather stations within both the 4- and 12-km gridded domains (see Fig. 2) for both day 1 and day 2 forecasts. To ensure a fair evaluation of HFWI values in relation to the observation-derived DFWI values—representing midafternoon [defined by Van Wagner (1987) as 1600 local time] fire danger calculated from standard noon local time weather observations—we compared the HFWI values at standard noon local, midafternoon local (1600 local time), and the maximum values occurring between midnight and midnight local time from the HFWI method against the observation-derived DFWI.
The choice of standard noon local was to assess how well the HFWI predicts fire behavior at the same time as the traditionally used weather observations. Midafternoon local HFWI values were selected to represent the assumed time of peak fire activity in the DFWI system. Finally, the decision to use the maximum HFWI values aligns with the purpose of the DFWI system, as it is meant to represent peak fire activity (Van Wagner 1987; Lawson and Armitage 2008). Considering all these HFWI values helps us to assess HFWI behavior in relation to the DFWI.
-
The HFWI method relies on the weather-sensitive FFMC, which is computed hourly using different empirical formulas than for the FFMC from the daily method (Van Wagner 1977, 1987). The daily FFMC assumes a diurnal cycle which does not necessarily capture the true fluctuations in hourly weather conditions (Van Wagner 1977) for our study region. Consequently, the daily and hourly FFMC values are not expected to align perfectly due to the inherent variability in hour-to-hour weather conditions and differing equations. This variability subsequently extends to the DFWI and HFWI calculations. However, it is important to compare the noon, midafternoon, and maximum HFWI values, as it maintains a relationship with the traditional daily method, which is commonly used in many regions across the world (Van Wagner 1987; Taylor and Alexander 2006; Cruz and Plucinski 2007; Mölders 2010; De Groot et al. 2015; Field et al. 2015; Di Giuseppe et al. 2016; Vitolo et al. 2020; Mandal et al. 2022).
-
Furthermore, the FWI system relies solely on weather inputs, meaning that the accuracy of the DFWI and HFWI methods is contingent on the quality of the underlying NWP model(s) used in their calculations.
The Taylor diagram comparing the forecasted noon, midafternoon, and maximum HFWI and the DFWI against the observation-derived DFWI values (Fig. 8) indicates that the forecasted DFWI outperforms the forecasted noon, midafternoon, and maximum HFWI values when verified against observed DFWI across the varied statistics. This is because the observation-derived DFWI values use the same empirical formulas as the DFWI forecast, whereas HFWI variables are derived from different formulas. However, it is noteworthy that the noon, midafternoon, and maximum HFWI forecast values exhibit comparable correlation with the observation-derived DFWI values as the DFWI forecast values.
Taylor diagram showing FWI forecast performance statistics for daily (circles), hourly noon (squares), hourly midafternoon (pentagon), and hourly maximum (diamonds) values, for day 1 and day 2 forecasts. Forecast bias is also shown. Red (blue) symbols denote the 4-km (12-km) domain. The black triangle location indicates a perfect forecast (observation baseline).
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
The biases associated with the noon, midafternoon, and maximum HFWI forecasts are expected in the context of how they would compare to the observation-derived DFWI values. The noon HFWI values have a larger negative bias because standard noon local time usually does not coincide with the timing of peak fire behavior.
The midafternoon HFWI values are closest to the observed DFWI values when compared to the noon and maximum HFWI values. However, they do not align with the observed DFWI values as well as the forecasted DFWI values due to the inherent variability in hour-to-hour weather conditions and differing equations.
The maximum HFWI values are slightly higher than the observed DFWI values, as shown by the std dev (Fig. 8). This suggests that the HFWI approach can capture fire weather conditions that may contribute to more intense fire activity at times of the day other than the assumed midafternoon 1600 local time built into the DFWI system. This aspect is explored and substantiated in the following section through a comparison with satellite-observed FRP for nine wildfire case studies. This comparison underscores the potential utility of the HFWI method for wildfire agencies worldwide.
For DFWI and HFWI midafternoon predictions, the domain with finer 4-km grid cells has a performance advantage over the 12-km domain across all statistical metrics (see Table 1 to compare values). For the HFWI maximum forecasts, the 12-km WRF domain gives a slightly better forecast. To assess why, we looked at the forecast performance of the four weather inputs used to calculate the FWI system for the two domains, specifically Pearson’s correlation and MBE between the observed and modeled data (see appendix A for details).
Pearson’s correlation r, MBE, RMSD, and standard deviation difference (std dev diff.) between the observed and modeled HFWI and DFWI data at 917 weather stations from the day 1 forecast. The “best” metric value for each method is in bold.
b. Wildfire case studies analysis
To further verify and demonstrate the utility of the HFWI method, it was compared against FRP data observed from the ABI aboard the GOES-R satellites for nine different wildfire cases during the 2021, 2022, and 2023 wildfire seasons (Table 2). It is important to highlight that FWI and FRP represent two different aspects of wildfires: FWI primarily relies on weather conditions and fuel moisture content to predict potential future fire behavior, whereas FRP quantifies the thermal output of an active wildfire. Limitations may arise in this comparison from approximated or unaccounted factors within the FWI system, such as fuel type, topography, and fire suppression. The intent of this comparison is to provide a better understanding of the HFWI system performance with respect to predicting fire intensity.
Details of the nine wildfire case studies, including names, dates of occurrence, locations, Pearson’s correlations between HFWI and HFRP, the WRF domain, and the geostationary satellite used for comparison.
For the first case, the Caldor Fire (mentioned previously) ignited on 14 August 2021 within the Sierra Nevada range of eastern California. The analysis period spans from 16 to 25 August 2021, for which we compare a time series of the forecasted DFWI and HFWI with observed hourly FRP (see Fig. 9). The 2-day delay in the start of our analysis is due to an absence of processed FRP observations from the GOES-West satellite on 14–15 August 2021.
GOES-West HFRP (MW; red line) estimated by algorithm vs HFWI (solid blue line) and DFWI (dashed blue line) data, derived from the WRF 12-km domain using the first 24-h of spun-up daily forecast data, over the time period 16–25 Aug 2021, during the Caldor Fire near South Lake Tahoe, CA. Dotted red lines represent the periods of missing data, where endpoints are linearly interpolated (not included in statistical analysis). Also shown are Pearson’s correlation between HFWI and FRP values and the percentage of observed values out of the total possible observations for the case study analysis period.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
Between 16 and 18 August 2021, the observed FRP from the Caldor Fire exhibited higher values than during the rest of the case study analysis period. This increase in FRP values was likely a consequence of a synoptic ridge breakdown and the passage of a dry cold front, which increased wind speed over the fire at 2000 PDT 17 August 2021. These findings are supported by the analysis of ERA5 50-kPa geopotential heights and surface analysis archived maps from the National Weather Service (NWS) Weather Prediction Center (WPC) (Hersbach et al. 2023; NWS 2023). Following 18 August 2021, the observed FRP data began to follow an atypical diurnal cycle.
On the nights of 17, 19, and 20 August 2021, the wildfire exhibited surges in emitted radiative energy around midnight local time. While the HFWI method failed to predict a peak in fire danger on 17 August around midnight, likely due to forecasted wind speeds being weaker than those on subsequent nights, it did forecast peak fire danger to occur around midnight on the nights of 19 and 20 August. We attribute this successful prediction by HFWI to increased wind speeds and minimal nighttime recovery of relative humidity, as predicted by the NWP forecasts (see appendix B, Fig. B1; in appendix B, we explore the interplay of the four weather inputs with respect to forecasted DFWI and HFWI for the four case studies discussed in this section). This is an important result, since the DFWI, relying on weather conditions from standard noon local time, is unable to predict peak fire conditions during the evening and overnight (by definition).
Another case study is the Oak Fire near Mariposa, California, United States, in 2022 (Fig. 10). Over the three and a half days, the Oak Fire exhibited a gradual decrease in emitted heat energy, which the HFWI predicted. A key period of interest for the Oak Fire occurred from midday on 23 August to the early morning of 24 August local time, when the wildfire experienced a multihour period of sustained high FRP values, some of which occurred in the late evening (early morning) of 23 (24) August. While the HFWI model missed the onset of this period of high FRP by approximately 2 h, due to elevated 2-m relative humidity values and a lull in forecasted wind speeds, it did predict the prolonged period of high fire danger, which was predominantly a wind-driven event [see Fig. B2 (appendix B)].
As in Fig. 9, but for the time period 22–26 Jul 2022 during the Oak Fire near Mariposa, CA. The FWI data were derived from the WRF 12-km domain, and FRP data were observed by GOES-West.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
An additional case study example is the Wildcat Fire near Hudson Bay, Saskatchewan, Canada, in July 2021 (Fig. 11). The wildfire was ignited by a lightning strike on 14 July 2021 but remained undetected by the GOES algorithm until 22 July. On 23 July 2021, warm air advection and increasing wind speeds ahead of an approaching cold front led to prefrontal heating and drying of forest fuels. Upon frontal passage, wind speeds increased, resulting in a spike in HFWI at 1400 local time 23 July. After the cold frontal passage, air temperature cooled; dewpoint temperature increased; and winds weakened, which led to a gradual decline in predicted fire danger [see Fig. B3 (appendix B)]. This same pattern is reflected in the observed FRP and in the surface analysis maps from the NWS–WPC archive.
As in Fig. 9, but for the time period 22–26 Jul 2021, during the Wildcat Fire, Hudson Bay, SK. The FWI data were derived from the WRF 4-km domain, and FRP data were observed by GOES-West.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
The final case study that we discuss in detail in this paper, the Marshall Fire, occurred on 30 December 2021 in Boulder, Colorado, United States (Fig. 12). The fire burned within the wildland–urban interface (WUI) in a short grass fuel type, driven by strong downslope winds of ∼25 m s−1 (with gusts > 30 m s−1), across the front range of the Rocky Mountains (Fovell et al. 2022; Juliano et al. 2023). The fast winds drove rapid fire spread and destroyed 1091 buildings in less than 36 h (Fovell et al. 2022; Boulder County Office of Disaster Management 2022). Although the 12-km domain WRF model driving the HFWI underpredicted wind speed (by ∼7 m s−1), it did capture the timing of the event, resulting in high fire danger rating values [see Fig. B4 (appendix B)]. This is of particular note as the National Weather Service did not issue any red flag warning leading up to the event due to relative humidity values not dropping below the 15% criteria in this region [Columbia Broadcasting System (CBS) Colorado 2022].
As in Fig. 9, but for the time period 30–31 Dec 2021, during the Marshall Fire, Boulder, CO. The FWI data were derived from the WRF 12-km domain, and FRP data were observed by GOES-East.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
Table 2 provides the name, date, location, and Pearson’s correlation for all nine case studies. Upon thorough analysis of all the wildfire cases, a clear pattern emerged: while both the HFWI and DFWI methods exhibit strong correlations with observed FRP values, neither method can accurately capture the magnitude of these observed FRP values. This limitation arises from the nature of the FWI system, which is designed to be fuel-type independent and relies solely on weather conditions. Consequently, the FWI values at different locations may not be directly comparable in terms of FRP. Additionally, various spatial locations possess distinct climatological characteristics, making it evident that a specific FWI value does not equate to the same level of potential fire danger across all locations.
One potential solution to address this issue is to spatially normalize the system. We are currently working to establish an hourly and daily fire weather system climatology, using 30 years of ERA5-Land data, to enable spatial normalization. Preliminary results are promising, as they bring the magnitude of the normalized FWI values into closer alignment with the magnitudes of observed FRP values at various locations.
4. Conclusions
This study presents a comprehensive analysis of an hourly FWI (HFWI) forecasting system driven by the WRF NWP model, and its comparative performance with the traditional daily FWI (DFWI) approach. It also underscores the potential benefits of an HFWI system in delivering more accurate fire danger predictions, especially when fire behavior departs from the conventional midafternoon peak. Several key takeaways emerge from this investigation as follows:
a. HFWI versus DFWI
The HFWI method, while not expected to precisely match DFWI values due to the inherent variability in hourly weather conditions, exhibits a strong correlation with the observed DFWI values that fire agencies use in day-to-day operations. Also, the maximum values derived from the HFWI method, though slightly higher than DFWI observations, suggest its ability to predict atypical extreme fire weather conditions that deviate from the typical midafternoon occurrences. This includes instances when extreme fire weather conditions occur at night and early morning. Additionally, the HFWI approach indicates the presence of multiple local maxima in fire behavior within a single day when this occurs—a capability absent in the traditional daily system, by definition. This emphasizes the HFWI method’s proficiency in predicting diverse temporal patterns of heightened fire weather, contributing to a more comprehensive understanding of potential risk factors for extreme fire behavior.
b. Wildfire case studies
For nine wildfire case studies, the HFWI system demonstrates its capability to forecast shifts in fire danger timing from the assumed DFWI midafternoon peak, including periods when fire activity intensifies during the evening and nighttime. These cases showcase the ability of the system to capture the changes in fire danger associated with changing weather conditions (e.g., cold frontal passage and downslope winds).
c. Forecasting
The HFWI system holds promise for enhancing the precision of fire danger magnitude forecasts and offering more information to wildfire agencies on the timing of peak fire activity in a forecast day. The accuracy, however, is contingent on the accuracy of the underlying NWP driving it, as errors in the weather forecast will cause errors in predicting fire danger.
d. Open-source model
This HFWI system is open-source to encourage collaboration, enhancements, and adaptability to other geographical regions and NWP models.
This research highlights the value of the HFWI approach in improving fire danger assessments and predictions, thus enhancing wildfire management, especially during unconventional fire behavior. The study sets the stage for further HFWI integration into wildfire decision-making. Future work underway will attempt to normalize HFWI with respect to regional climatologies.
Acknowledgments.
We acknowledge the support of our funders: NSERC Discovery (RGPIN-2017-03849, April 2017–March 2023), NRCan (WFR2324-UBC, March 2023–April 2024), BC Ministry of Environment and Climate Change Strategy (GS23EPESB0117SF, March 2022–April 2023, and GS23EPESB0150, August 2023–April 2024), AB Ministry of Environment and Parks (23POL832, April 2023–November 2024), Government Northwest Territories (CA 4287, 2017–24), BC Hydro and Power Authority (00091424, 2016–24), MITACS, Inc. (IT28208, 2022–27), and NSERC Western Grid Resilience (ALLRP 585094 – 23, 2023–28). Their generous funding has been instrumental in advancing our research objectives. We also thank all members of the UBC Weather Forecast Research Team for their encouragement and feedback.
Data availability statement.
1) The forecasted hourly and daily fire weather index along with associated meteorological inputs predicted from the Weather Research and Forecasting (WRF) Model from 1 January 2021 to 31 December 2023 can be found at https://doi.org/10.20383/103.0876. 2) The fire weather forecast (FWF) model Python code, along with documentation detailing the configuration, execution, and working with the model data, is publicly available on GitHub under an open-source MIT license. The GitHub repository and documentation can be accessed at https://github.com/cerodell/fwf/tree/ams-fwi and https://cerodell.github.io/fwf-docs. 3) Current forecasts of FWF codes and indices, along with associated meteorological information and the latest wildfire locations, can be viewed on a zoomable map with pop-up point forecast functionality. The information is updated daily and is accessible at https://firesmoke.ca/forecasts/fireweather/. 4) Hourly reanalysis data (1 January–31 December 2020) from the Fifth major global reanalysis produced by European Centre for Medium-Range Weather Forecasts (ECMWF) (ERA5) model was downloaded from the Copernicus Climate Data Store at https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels. 5) The fire radiative power (FRP) data, downloaded from the National Oceanic and Atmospheric Administration (NOAA) Geostationary Operational Environmental Satellites (GOESs), GOES-16, GOES-17, and GOES-18, were accessed on 10 Jan 2023 from https://registry.opendata.aws/noaa-goes. 6) The surface fire weather station observation data, downloaded from the Natural Resources Canada (NRCan) Canadian Wildland Fire Information System (CWFIS), were accessed on 5 May 2023 from https://cwfis.cfs.nrcan.gc.ca/downloads/fwi_obs/. 7) The Taylor diagrams were created using SkillMetrics (Rochford 2016). 8) The FWF model made use of the xarray Python package (Hoyer and Hamman 2017).
APPENDIX A
Performance of 4- and 12-km WRF Domains for FWI Predictions
This appendix compares the performance of the WRF 4- and 12-km grid cell domains in predicting the fire weather index (FWI), by analyzing the four weather inputs for both domains using Pearson’s correlation and mean bias error (MBE). Precipitation statistics were computed for events where observed precipitation exceeded a 0.1-mm threshold over a 24-h period. We used the standard noon local observations to align with those used for the observation-derived DFWI.
The results indicate that both domains performed nearly equally according to Pearson’s correlation, for each of the four weather inputs (Table A1). However, the MBE for 2-m temperature and 2-m relative humidity reveals a cool, more saturated bias in both domains, which is more pronounced in the 12-km domain compared to the 4-km domain (Fig. A1). We attribute this discrepancy to the finer 4-km grid spacing providing a more accurate representation of the complex terrain and its effects within the study area.
Pearson’s correlation r and MBE values for 4- and 12-km WRF domain predictions of FWI weather input variables at 917 weather stations: 2-m temperature, 2-m RH, 10-m wind speed, and 24-h precipitation. The best metric value for each variable is in bold.
Scatterplot showing true weather station elevation (y axis; m) vs nearest model grid cell elevation to weather station location (x axis; m) for the (left) 4- and (right) 12-km WRF domains. Colors represent the MBE for (a),(b) 2-m temperature, (c),(d) 2-m RH, (e),(f) 10-m wind speed, and (g),(h) 24-h precipitation. Each panel subplot contains a histogram showing the distribution of the MBE for its respective weather variable.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
Specifically, in Fig. A1, the elevation of the 4-km domain (Figs. A1a,c,e, and g) is more aligned with the true elevation of weather stations compared to the 12-km domain (Figs. A1b,d,f, and h). Analyzing 2-m temperature and 2-m relative humidity (Figs. A1a–d) reveals that when the model terrain is higher (lower) than the true weather station elevation, it leads to a cool and more saturated (warm and less saturated) bias, as expected. Notably, more weather stations were situated in lower elevation areas (i.e., valleys) than high elevation areas (i.e., mountain tops), leading to a more pronounced cool and more saturated bias.
Concerning precipitation, both domains exhibit a positive bias, which is slightly smaller for the 4-km domain, with little evidence of elevation bias affecting this (Figs. A1e,f).
For 10-m wind speed, a positive bias (overprediction of wind speed) was observed, again with little evidence of elevation bias affecting the wind speed bias (Figs. A1g,h). The overprediction of wind speed compensates for the overall cool and moist bias with respect to FWI predictions, with the 4-km domain exhibiting a higher positive bias than the 12-km domain. This compensation is more pronounced because, in the absence of precipitation, the FWI system is most sensitive to 10-m wind speed, followed by 2-m relative humidity and 2-m temperature (Dowdy et al. 2010).
Due to the sensitivity of the FWI system, the more pronounced overprediction of wind in the 4-km domain, combined with the better representation of complex terrain, reduced the cool and more saturated bias more than in the 12-km domain, resulting in improved FWI predictions for the finer 4-km domain overall.
APPENDIX B
Interplay of Weather Inputs for FWI Predictions
In this appendix, we illustrate the interplay of the four weather inputs with respect to forecasted DFWI and HFWI systems for the four case studies discussed in detail in results and discussion (section 3. For each case, we generated a time series plot for the case study dates, depicting WRF-forecasted weather inputs and related indices and codes from the DFWI and HFWI methods. Specifically, the time series for each case were constructed using the FRP-weighted spatially average method, as detailed in section 2.
As an example, Fig. B1 illustrates the hourly FFMC, ISI, and FWI (solid lines), as well as the daily FFMC, DMC, DC, ISI, and FWI (diamonds) for the Caldor Fire case study. In the lower part of Fig. B1, the hourly (solid lines) and daily (diamonds) weather inputs are depicted. The DFWI system codes and indices (diamonds) are placed at 1600 local time based on the underlying assumption embedded in the DFWI system (Van Wagner 1987; Lawson and Armitage 2008). The daily weather inputs (diamonds) correspond to standard noon local time, which the daily system utilizes to predict fire activity during the assumed midafternoon period (1600 local time).
Time series of (bottom) WRF-forecasted weather inputs (blue = 2-m RH, red = 2-m temperature, black = 10-m wind speed, and green = 24-h precipitation) and (top) FWI systems indices and codes (brown = DC, green = DMC, orange = FFMC, purple = ISI, gray = BUI, and red = FWI) for 16–25 Aug 2021. All times are in local standard time for the Caldor Fire near South Lake Tahoe, CA. Solid lines represent the hourly weather and HFWI forecast data. Diamonds represent standard noon local time weather and 1600 local time DFWI forecast data. Vertical grid lines represent midnight local time. Weather inputs and FWI data are generated and derived from the WRF 12-km domain.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
In the Caldor Fire case study (section 3), the HFWI accurately predicts peaks in fire danger (peaks in FWI) around midnight on the nights of 19 and 20 August, in comparison to HFRP as shown in Fig. 9. This successful prediction by HFWI is attributed to increased wind speeds and minimal nighttime recovery of relative humidity, as forecasted by WRF.
The second example is the Oak Fire near Mariposa, California, United States, in 2022 (Figs. 10 and B2). Over the three and a half days, the Oak Fire exhibited a gradual decrease in emitted heat energy, which the HFWI predicted. A key period of interest for the Oak Fire occurred from midday on 23 August to the early morning of 24 August local time when the wildfire experienced a multihour period of sustained high FRP values, some of which occurred in the late evening (early morning) of 23 (24) August. While the HFWI model missed the onset of this period of high FRP by approximately 2 h, it did predict the prolonged period of high fire danger, which was predominantly a wind-driven event.
As in Fig. B1, but for the time period 22–26 Jul 2022 during the Oak Fire near Mariposa, CA. Vertical grid lines are every 12 h (midday and midnight) local time. Weather inputs and FWI data are generated and derived from the WRF 12-km domain.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
The third case study example is the Wildcat Fire near Hudson Bay, Saskatchewan, Canada, in July 2021 (Figs. 11 and B3). On 23 July 2021, warm air advection and increasing wind speeds ahead of an approaching cold front led to prefrontal heating and drying of forest fuels. Upon frontal passage, wind speeds increased, resulting in a spike in HFWI at 1400 local time 23 July. After the cold frontal passage, air temperature cooled; dewpoint temperature increased; and winds weakened, which led to a gradual decline in predicted fire danger.
As in Fig. B1, but for the time period 22–26 Jul 2021, during the Wildcat Fire, Hudson Bay, SK. Vertical grid lines are every 12 h (midday and midnight) local time. Weather inputs and FWI data are generated and derived from the WRF 4-km domain.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
The final case study that we discussed in detail in section 3, the Marshall Fire, occurred on 30 December 2021 in Boulder, Colorado, United States (Figs. 12 and B4). The fire burned within the wildland–urban interface (WUI) in a short grass fuel type, driven by strong downslope winds of ∼25 m s−1 (with gusts > 30 m s−1), across the front range of the Rocky Mountains (Fovell et al. 2022; Juliano et al. 2023). The fast winds drove rapid fire spread and destroyed 1091 buildings in less than 36 h (Fovell et al. 2022; Boulder County Office of Disaster Management 2022). Although the 12-km domain WRF model driving the HFWI underpredicted wind speed (by ∼7 m s−1), it did capture the timing of the event, resulting in high fire danger rating values.
As in Fig. B1, but for the time period 30–31 Dec 2021, during the Marshall Fire, Boulder, CO. Vertical grid lines are every 4 h local time. Weather inputs and FWI data are generated and derived from the WRF 12-km domain.
Citation: Weather and Forecasting 39, 6; 10.1175/WAF-D-23-0226.1
REFERENCES
Andrews, P. L., D. O. Loftsgaarden, and L. S. Bradshaw, 2003: Evaluation of fire danger rating indexes using logistic regression and percentile analysis. Int. J. Wildland Fire, 12, 213–226, https://doi.org/10.1071/WF02059.
Balch, J. K., J. T. Abatzoglou, M. B. Joseph, M. J. Koontz, A. L. Mahood, J. McGlinchy, M. E. Cattau, and A. P. Williams, 2022: Warming weakens the night-time barrier to global fire. Nature, 602, 442–448, https://doi.org/10.1038/s41586-021-04325-1.
Beck, J. A., and O. B. Armitage, 2004: Diurnal fine fuel moisture characteristics at a Northern Latitude. Proc. 22nd Tall Timbers Fire Ecology Conf. Fire in Temperate, Boreal, and Montane Ecosystems, Tallahassee, FL, Tall Timbers Research Station, 211–221, https://talltimbers.org/wp-content/uploads/2018/09/211-BeckandArmitage2004_op.pdf.
Bedia, J., S. Herrera, J. M. Gutiérrez, G. Zavala, I. R. Urbieta, and J. M. Moreno, 2012: Sensitivity of fire weather index to different reanalysis products in the Iberian Peninsula. Nat. Hazards Earth Syst. Sci., 12, 699–708, https://doi.org/10.5194/nhess-12-699-2012.
Boulder County Office of Disaster Management, 2022: Marshall Fire damage assessment. Boulder County Office of Disaster Management, accessed 27 October 2023, https://www.arcgis.com/apps/webappviewer/index.html?id=9f3314c39ad64fac925101aae0bdd62c.
Canadian Forest Service Fire Danger Group, 2021: An overview of the next generation of the Canadian forest fire danger rating system. Information Rep. GLC-X-26, 63 pp., https://cfs.nrcan.gc.ca/publications?id=40474.
Chui, T. C. Y., D. Siuta, G. West, H. Modzelewski, R. Schigas, and R. Stull, 2019: On producing reliable and affordable numerical weather forecasts on public cloud-computing infrastructure. J. Atmos. Oceanic Technol., 36, 491–509, https://doi.org/10.1175/JTECH-D-18-0142.1.
CBS Colorado, 2022: Experts explain why there was no red flag warning issued before Marshall Fire. Columbia Broadcasting System (CBS) Broadcasting Inc., accessed 10 October 2023, https://www.cbsnews.com/colorado/news/marshall-fire-red-flag-warning/.
Cruz, M. G., and M. P. Plucinski, 2007: Billo road fire: Report on fire behaviour phenomena and suppression activities. Bushfire Cooperative Research Centre Tech. Rep. A.07.02, 102 pp., https://www.bushfirecrc.com/sites/default/files/managed/resource/billoroadfirefinalreportweb_0.pdf.
de Groot, W. J., B. M. Wotton, and M. D. Flannigan, 2015: Wildland fire danger rating and early warning systems. Wildfire Hazards, Risks and Disasters, J. F. Shroder and D. Paton, Eds., Elsevier, 207–228., https://doi.org/10.1016/B978-0-12-410434-1.00011-7.
Di Giuseppe, F., F. Pappenberger, F. Wetterhall, B. Krzeminski, A. Camia, G. Libertá, and J. San Miguel, 2016: The Potential predictability of fire danger provided by numerical weather prediction. J. Appl. Meteor. Climatol., 55, 2469–2491, https://doi.org/10.1175/JAMC-D-15-0297.1.
Di Giuseppe, F., S. Rémy, F. Pappenberger, and F. Wetterhall, 2017: Improving forecasts of biomass burning emissions with the fire weather index. J. Appl. Meteor. Climatol., 56, 2789–2799, https://doi.org/10.1175/JAMC-D-16-0405.1.
Di Giuseppe, F., S. Rémy, F. Pappenberger, and F. Wetterhall, 2018: Using the Fire Weather Index (FWI) to improve the estimation of fire emissions from fire radiative power (FRP) observations. Atmos. Chem. Phys., 18, 5359–5370, https://doi.org/10.5194/acp-18-5359-2018.
Dowdy, A. J., G. A. Mills, K. Finkele, and W. de Groot, 2010: Index sensitivity analysis applied to the Canadian Forest Fire Weather Index and the McArthur Forest Fire Danger Index. Meteor. Appl., 17, 298–312, https://doi.org/10.1002/met.170.
Field, R. D., and Coauthors, 2015: Development of a Global Fire Weather Database. Nat. Hazards Earth Syst. Sci., 15, 1407–1423, https://doi.org/10.5194/nhess-15-1407-2015.
Flannigan, M. D., B. M. Wotton, G. A. Marshall, W. J. de Groot, J. Johnston, N. Jurko, and A. S. Cantin, 2016: Fuel moisture sensitivity to temperature and precipitation: Climate change implications. Climatic Change, 134, 59–71, https://doi.org/10.1007/s10584-015-1521-0.
Forestry Canada, 1992: Development and structure of the Canadian forest fire behavior prediction system. Forestry Canada, Headquarters, Fire Danger Group and Science and Sustainable Development Directorate Information Rep. ST-X-3, 64 pp., https://cfs.nrcan.gc.ca/publications?id=10068.
Fovell, R. G., M. J. Brewer, and R. J. Garmong, 2022: The December 2021 Marshall Fire: Predictability and gust forecasts from operational models. Atmosphere, 13, 765, https://doi.org/10.3390/atmos13050765.
Freeborn, P. H., W. M. Jolly, M. A. Cochrane, and G. Roberts, 2022: Large wildfire driven increases in nighttime fire activity observed across CONUS from 2003–2020. Remote Sens. Environ., 268, 112777, https://doi.org/10.1016/j.rse.2021.112777.
Groot, W. J., R. D. Field, M. A. Brady, O. Roswintiarti, and M. Mohamad, 2007: Development of the Indonesian and Malaysian Fire Danger Rating Systems. Mitigation Adapt. Strategies Global Change, 12, 165–180, https://doi.org/10.1007/s11027-006-9043-8.
Hersbach, H., 2023: ERA5 hourly data on single levels from 1940 to present. Copernicus Climate Change Service (C3S) Climate Data Store (CDS), accessed 31 March 2023, https://doi.org/10.24381/cds.adbb2d47.
Hersbach, H., and Coauthors, 2023: ERA5 hourly data on pressure levels from 1940 to present. Copernicus Climate Change Service (C3S) Climate Data Store (CDS), accessed 31 March 2023, https://doi.org/10.24381/cds.bd0915c6.
Higuera, P. E., M. C. Cook, J. K. Balch, E. N. Stavros, A. L. Mahood, and L. A. St. Denis, 2023: Shifting social-ecological fire regimes explain increasing structure loss from western wildfires. PNAS Nexus, 2, pgad005, https://doi.org/10.1093/pnasnexus/pgad005.
Hoyer, S., and J. Hamman, 2017: Xarray: N-D labeled arrays and datasets in Python. J. Open Res. Software, 5, 10, https://doi.org/10.5334/jors.148.
Jolly, W. M., M. A. Cochrane, P. H. Freeborn, Z. A. Holden, T. J. Brown, G. J. Williamson, and D. M. J. S. Bowman, 2015: Climate-induced variations in global wildfire danger from 1979 to 2013. Nat. Commun., 6, 7537, https://doi.org/10.1038/ncomms8537.
Juliano, T. W., and Coauthors, 2023: Toward a better understanding of wildfire behavior in the wildland-urban interface: A case study of the 2021 Marshall Fire. Geophys. Res. Lett., 50, e2022GL101557, https://doi.org/10.1029/2022GL101557.
Lawson, B. D., and O. B. Armitage, 2008: Weather guide for the Canadian forest fire danger rating system. Natural Resources Canada, Canadian forest Service, Northern Forestry Centre Tech. Rep., 87 pp., https://cfs.nrcan.gc.ca/publications?id=29152.
Lelieveld, J., J. S. Evans, M. Fnais, D. Giannadaki, and A. Pozzer, 2015: The contribution of outdoor air pollution sources to premature mortality on a global scale. Nature, 525, 367–371, https://doi.org/10.1038/nature15371.
Li, F., X. Zhang, S. Kondragunta, C. C. Schmidt, and C. D. Holmes, 2020: A preliminary evaluation of GOES-16 active fire product using Landsat-8 and VIIRS active fire data, and ground-based prescribed fire records. Remote Sens. Environ., 237, 111600, https://doi.org/10.1016/j.rse.2019.111600.
Mandal, A., G. Nykiel, T. Strzyzewski, A. Kochanski, W. Wrońska, M. Gruszczynska, and M. Figurski, 2022: High-resolution fire danger forecast for Poland based on the Weather Research and Forecasting Model. Int. J. Wildland Fire, 31, 149–162, https://doi.org/10.1071/WF21106.
Mölders, N., 2010: Comparison of Canadian Forest Fire Danger Rating System and National Fire Danger Rating System fire indices derived from Weather Research and Forecasting (WRF) model data for the June 2005 interior Alaska wildfires. Atmos. Res., 95, 290–306, https://doi.org/10.1016/j.atmosres.2009.03.010.
National Centers for Environmental Prediction, National Weather Service, NOAA, U.S. Department of Commerce, 2015: National Centers for Environmental Prediction North American Mesoscale (NAM) 12 km analysis. National Center for Atmospheric Research, Computational and Information Systems Laboratory, accessed 1 January 2021, https://doi.org/10.5065/G4RC-1N91.
Natural Resources Canada, Canadian Forest Service, 2023: Canadian forest fire weather index system data from the Canadian wildland fire information system, version 3.0. Northern Forestry Centre, accessed 31 October 2023, http://cwfis.cfs.nrcan.gc.ca/downloads/fwi_obs/.
Natural Resources Canada Canadian Wildland Fire Information System, 2023: Data sources and methods: Canadian Forest Fire Weather Index (FWI) System. Natural Resources Canada, accessed 31 October 2023, https://cwfis.cfs.nrcan.gc.ca/background/dsm/fwi.
NWS, 2023: Surface analysis archive maps. National Weather Service Weather Prediction Center, accessed 10 September 2023, https://www.wpc.ncep.noaa.gov/html/sfc-zoom.php.
Rochford, P. A., 2016: SkillMetrics: A Python package for calculating the skill of model predictions against observations. GitHub, accessed 11 September 2020, http://github.com/PeterRochford/SkillMetrics.
Schmidt, C., J. Hoffman, E. Prins, and S. Lindstrom, 2012: GOES-R Advanced Baseline Imager (ABI) algorithm theoretical basis document for fire/hot spot characterization. NOAA Tech. Rep.NESDIS STAR, 106 pp., https://www.star.nesdis.noaa.gov/goesr/docs/ATBD/Fire.pdf.
Skamarock, W. C., and Coauthors, 2020: A description of the Advanced Research WRF model version 4.3. NCAR Tech. Rep. NCAR/TN-556+STR, 165 pp., https://doi.org/10.5065/1dfh-6p97.
Taylor, S. W., and M. E. Alexander, 2006: Science, technology, and human factors in fire danger rating: The Canadian experience. Int. J. Wildland Fire, 15, 121–135, https://doi.org/10.1071/WF05021.
Van Wagner, C. E., 1974: Structure of the Canadian forest fire weather index. Canadian Forestry Service Tech. Rep. 1333, 49 pp., https://cfs.nrcan.gc.ca/publications?id=24864.
Van Wagner, C. E., 1977: A method of computing fine fuel moisture content throughout the diurnal cycle. Canadian Forestry Service Tech. Rep. PS-X-69, 17 pp., https://cfs.nrcan.gc.ca/publications?id=25591.
Van Wagner, C. E., 1987: Development and structure of the Canadian Forest Fire Weather Index System. Canadian Forestry Service Tech. Rep. 35, 48 pp., https://cfs.nrcan.gc.ca/publications?id=19927.
Van Wagner, C. E., and T. L. Pickett, 1985: Equations and FORTRAN program for the Canadian forest fire weather index system. Canadian Forestry Service Tech. Rep. 33, 25 pp., https://cfs.nrcan.gc.ca/publications?id=19973.
Vitolo, C., F. Di Giuseppe, C. Barnard, R. Coughlan, J. San-Miguel-Ayanz, G. Libert, and B. Krzeminski, 2020: ERA5-based global meteorological wildfire danger maps. Sci. Data, 7, 216, https://doi.org/10.1038/s41597-020-0554-z.
Wang, X., B. M. Wotton, A. S. Cantin, M.-A. Parisien, K. Anderson, B. Moore, and M. D. Flannigan, 2017: cffdrs: An R package for the Canadian Forest Fire Danger Rating System. Ecol. Processes, 6, 5, https://doi.org/10.1186/s13717-017-0070-z.
Wooster, M. J., G. Roberts, G. L. W. Perry, and Y. J. Kaufman, 2005: Retrieval of biomass combustion rates and totals from fire radiative power observations: FRP derivation and calibration relationships between biomass consumption and fire radiative energy release. J. Geophys. Res., 110, D24311, https://doi.org/10.1029/2005JD006318.
