Wildfire Pyroconvection and CAPE: Buoyancy’s Drying and Atmospheric Intensification—Fort McMurray

Abstract

The accurate prediction of wildfire behavior and spread is possible only when fire and atmosphere simulations are coupled. In this work, we present a mechanism that causes a small fire to intensify by altering the atmosphere. These alterations are caused by fire-related fluxes at the surface. The fire plume and fluxes increase the convective available potential energy (CAPE) and the chance of the development of a strong pyroconvection system. To study this possible mechanism, we used WRF-Fire to capture fire line propagation as the result of interactions between heat and moisture fluxes, pressure perturbations, wind shear development and dry air downdraft. The wind patterns and dynamics of the pyroconvection system are simulated for the Horse River wildfire at Fort McMurray, Canada. The results revealed that the updraft speed reached up to 12 m/s. The entrainment mixed the mid and upper-level dry air and lowered the atmospheric moisture. The mid-level and upper-level dew point temperature changed by 5–10 ${}^{\circ }$C in a short period of time. The buoyant air strengthened the ascent as soon as the nocturnal inversion was eliminated by daytime heating. The 887 J/kg total increase of CAPE in less than 5 h and the high bulk Richardson number (BRN) of 93 were indicators of the growing pyro-cumulus cell. The presented simulation has not improved the original model or supported leading-edge numerical weather prediction (NWP) achievements, except for adapting WRF-Fire for Canadian biomass fuel. However, we were able to present a great deal of improvements in wildfire nowcasting and short-term forecasting to save lives and costs associated with wildfires. The simulation is sufficiently fast and efficient to be considered for a real-time operational model. While the project was designed and succeeded as an NWP application, we are still searching for a solution for the intractable problems associated with political borders and the current liable authorities for the further development of a new generation of national atmosphere–wildfire forecasting systems.

1. Introduction

Climate and weather impact the frequency and behavior of wildfires [1,2]. Wildfires also provide substantial short and long-term feedbacks to the atmosphere and land-surface [3]. The burning of biomass increases the output of emission products, heat and water vapor fluxes into the atmosphere; the short-term impact on the atmosphere is an immediate small-scale fire caused by local heat and moisture fluxes such as fire winds and pyro-cumulus clouds. Long-term atmospheric cooling can be caused by indirect and direct atmospheric aerosol radiative scattering or warming caused by released greenhouse gases such as carbon monoxide, methane and nitrous oxide [4,5,6,7]. Every year, global fires directly emit 200−300 Gkg($={10}^{15}$g) of carbon into the atmosphere [7]. Emissions from Canadian boreal biomass burning have regularly been observed in the upper troposphere and even in the lower stratosphere over Europe and the USA [8]. Our capacity to understand and manage wildfires will remain imperfect as long as our knowledge is confined by our limited observation and restricted understanding of wildfire mechanisms. One of the areas that requires investigation is the transition of a wildfire to a large fire with a high intensity, spread and burned area. The current study offers a numerical simulation case-study and discussion of the dynamics and physics of the transition of wildfire to pyroconvection. The research also suggests that a coupled atmosphere–fire prediction system could provide additional guidance for decision-makers.

Certain atmospheric conditions can result in excessive fuel drying and extreme pyroconvection activities [9]. Moisture in the mid-atmosphere tends to be resilient to rapid changes in the absence of dry/moist large-scale advection [10,11], while wildfire can change the atmosphere moisture content rapidly [12]. A lack of wildfire–atmosphere heat and moisture interaction can throw off any non-coupled weather or wildfire forecasts.

The wildfires associated with pyroconvection behavior cannot be predicted by a simple two-dimensional physical or semi-physical spread model. Although the simplified fire spread models are a useful measure of fire front spread on the ground in the absence of the forceful and rapid changes in the wind and moisture of the atmosphere boundary layer, they cannot foresee the time and location of occurrence of pyroconvection and extreme forces.

Primary knowledge [13] of mesoscale convection systems (MCSs) suggests three requirements for the development of deep convections: (1) instability, (2) an abundance of boundary layer moisture and (3) a lifting mechanism (trigger) to produce vertical motion, allowing a parcel to attain free convection. The requirements are fundamental to the development of all MCSs (or pyroconvection mesoscale convection systems, PMCSs) from short-lived single cells to destructive supercells. What separates the environments of varying storm types from each other is the vertical wind shear within the pre-storm environment. Buoyancy primarily determines the updraft strength, and wind shear determines the storm evolution [13]. Wildfire surface heating can act as a trigger, allowing free convection and adding moisture (via fuel combustion) to the air in a short period of time, which will increase buoyancy. Wildfires can modify the pre-existing environment toward a favorable deep convection condition by warming and moistening the lowest level of the stratosphere.

The pyroconvection grows with very strong winds which change directions and drive with lateral vorticity and vertical wind shear. Badlan et al. [14] performed an idealized numerical simulation to study the role of deep flaming in pyroconvection by using WRF-Fire version 3.9. They repeated the simulation with variations of fire intensities, spatial extensions and horizontal resolutions and three different atmospheric profiles to explore the dynamics of pyroconvection. These simplified simulations gave an informative insight into the dynamics of pyroconvection without considering the moisture fluxes. Lareau and Clements [12] showed the atmospheric vertical thermodynamics and dynamics profile, and the daily variation had a significant impact on the onset and development of pyroconvection, as the wildfire moisture and heat fluxes from below were shown to be able to expedite the process. They established direct observations using a scanning Doppler light detection and ranging (LiDAR) and mobile radiosonde system during two large wildfires in northern California. The results suggested that the plume condensation level was substantially higher than the ambient lifting condensation level (LCL), which implies that the lifting air in the fire plume reflects the actual properties of the lower plume. The plume must be cooled largely near its base due to entrainment as it decays towards adiabatic ascent higher up the convective condensation level (CCL).

To examine the transition process of a wildfire to a pyroconvection, we chose to study the early stages of the Horse River wildfire which eventually burned an area of 589,522 hectares. The emissions from the Horse River fire had a pronounced impact on the measured annual concentrations and health monitoring sites (NH₃ was up to 14 times higher than non-fire impacted periods for some pollutants) [6].

We are interested in the first days of the event because that is when the fire made the first transition to pyroconvection. We simulated the fire in hindcast mode using the numerical weather and fire prediction model. The choice of model depends on the questions at hand and the availability of the model and resources. After some investigations, we used WRF-Fire [15].

The objective of this paper is to investigate the transition of a wildfire to a large wildfire and the initiation of pyroconvection. We also investigate the practicality of predicting a large fire. We study the transition mechanism by using WRF 3.9.1. We present a Horse River wildfire (Fort McMurray, Alberta Canada) hindcast simulation, as well as our results, with a focus on atmosphere profiles and storm onset. The current research is not an observational or numerical advancement to wildfire models; it is rather an application and a study on the practicality and efficacy of wildfire forecasts. Lastly, we discuss the existing challenges to a fully functional operational coupled atmospheric–wildfire forecast system.

2. Horse River Wildfire 2016

The 2015–16 El Niño (NOAA El Niño Advisories) caused a warmer winter over Canada, at almost 5 ${}^{\circ }$C warmer than the climatological average across all provinces. The 2016 spring remained warmer and drier than the normal climate in western Canada and the Prairies, while an inflow of cold Arctic air brought colder than normal temperatures over eastern and northeastern Canada. The average precipitation for the period 1981–2010 for Fort McMurray in July was 80.7 mm and the maximum temperature was 24 ${}^{\circ }$C. Fort McMurray international airport weather station recorded winter temperatures which were about 4 ${}^{\circ }$C warmer than normal in 2015–2016, while precipitation was around half that of the normal records. This led to an extremely dry biomass condition combined with gusty winds at the beginning of May, creating critical conditions for wildfire with extreme fire danger indices in the area on May 1, 2016. The subject area is located in the northeast of Alberta, south-west of Fort McMurray. The city is in the middle of the Athabasca oil sands, surrounded by boreal forest of white spruce, jack pine, trembling aspen and peatlands. The overall topography of the area is relatively flat, while the four rivers that pass through the city of Fort McMurray create slopes with riparian vegetation of white spruce and balsam fir.

A severe wildfire in May 2016 led to the largest evacuation in the history of Canada and caused widespread damage. There were five active fires between April 29 and May 1, 2016 in the area. Four of them were within close distances from the populated area of the city of Fort McMurray; those fires (MMD001-004) were suppressed and controlled by Alberta Fire crews less than a day after their ignitions. The Horse River fire (identified by MWF009-16) started in the south of the city and swept through the city in less than 72 h. It destroyed over 2400 homes and buildings. It continued to spread across Alberta and into Saskatchewan. The Horse River fire started on May 1, 2016, and a mandatory evacuation was issued on May 3. The fire spread reached 30 to 40 m per minute and created fire-driven weather patterns and winds at its peak. The satellite image from the second day of the fire displayed the pyro-cumulus clouds. The burned area stretched west and east towards the Saskatchewan border, covering 589,552 hectares. The wildfire was declared to be under control on 4 July 2016 after burning for 65 days. During the fire, the lack of knowledge regarding the wind direction (due to the coarse observation and weather forecast biases) made it difficult to predict the fire behavior. The forecast uncertainty was large, which was partially due to fire wind modification and associated heat fluxes. The scientific questions regarding the fire are the following: What was the main trigger that unleashed the Horse River fire (MWF009-16)? What was the main difference between the Horse River fire and the other four fires in the area that had been suppressed? and could we have predicted the Horse River wildfire in a timely manner? To answer these questions, we examine the weather and fire as one system.

3. Methods

The WRF model is an upgrade of the fifth-generation NCAR/Penn State Mesoscale Model (MM5), and the fire module runs in large eddy simulation (LES) mode [16,17]. While WRF-Fire has been launched as a research and operational model in the USA and other countries [15], we could not make use of it in Canada. The problem is that the current Canadian biomass fuel classification is not compatible with the original WRF-Fire biomass classes. With some modifications and assumptions, we adjusted the Canadian digital biomass fuel classification map to be compatible with WRF-Fire and its fire module. Appendix A describes the technical modification of WRF-Fire for the Canadian digital biomass fuel map. The Horse River wildfire was simulated using a two-way interaction of the fire and the atmosphere. The case study was also fast and efficient computationally for real-time operational forecasts. Next, we briefly explain the case study area, data, the Canadian digital fuel map and finally the result and discussion of the case study. The Supplementary Materials provide the technical modifications of WRF-Fire for the Canadian digital biomass fuel map and the case study domains in detail.

3.1. Data

Initial and boundary conditions are required to start any environmental simulation. For the weather prediction/hindcast in a limited area, we used GFS (Global Forecast System by the National Centers for Environmental Prediction (NCEP)) analysis, which is publicly available. The resolution of GFS is very coarse for the fire spread algorithm in WRF-Fire, where the fire module requires a horizontal sub-grid resolution of 35 m${}^2$ or lower. To achieve such a fine resolution, we used the two-way nested simulation (multiple domains at different grid resolutions run simultaneously and communicate with each other). The coarser domain provided the initial and boundary conditions for the inside smaller domain (nest). The model ran multiple domains at multiple nested levels (telescoping) to produce very fine boundary conditions for the interior domain—i.e., a 200 m${}^2$ grid space resolution—where the fire ignited. We considered five two-way nested domains with 27,000, 9000, 3000, 1000 and 200 m${}^2$ grid space resolutions. To remain confined within the assumption that has been made by the fire-spread model [18], we defined the sub-grid resolution scale with a ratio of 0.1 of the inner WRF’s domain resolution (sub-grid resolution of 20 m${}^2$ from 200 m${}^2$ resolution of the inner WRF domain). Figure 1 shows the five WRF domains at the most inner domain. The inner domain presents the area in which the wildfire spread model interacts with the other weather domains.

In addition to weather data, the fire module requires fuel data and very fine resolution topography compatible with a sub-grid resolution. We added the digital elevation map from the Atlas of Canada (https://www.nrcan.gc.ca/home) with a horizontal resolution of 20 m${}^2$. Natural Resources of Canada provided us the available digital fuel map for Canada with a resolution of 250 m${}^2$. The Canadian digital fuel map has a slightly different classification for the variety of vegetation compared to the US digital map [19] which had been originally adopted in WRF-Fire (SFIRE). We modified the WRF-Fire “firenamelist” for the Canadian fuel classification (see the Supplementary Materials).

3.2. Observation

We used fire hotspots from FIRM (Fire Information for Resource Management System, https://earthdata.nasa.gov/firms) for ignitions, using MODIS (Moderate Resolution Imaging Spectroradiometer) Collection 6 NRT (near real-time, less than 3 h from the acquisition time) and VIIRS (Visible Infrared Imaging Radiometer Suite) hotspot data. The thermal anomalies/active fires represent any place within a 1 km pixel that is flagged by the MODIS Fire and Thermal Anomalies algorithm [20] as containing one or more fires within the pixel. This is the most basic fire product, in which active fires and other thermal anomalies, such as volcanoes, are identified as early as 3 h after acquisitions. The VIIRS Active Fire detection product is based on the instrument’s 375 m nominal resolution data. MODIS and VIIRS are multi-disciplinary sensors on-board NOAA and NASA Satellites (Terra, Aqua, Suomi-NPP). The earliest NRT (Figure 2 and Figure 3) hotspots in the Fort McMurray wildfire were detected by the Terra satellite with the MODIS instrument at 3:51 and 5:29 UTC (MDT+7) on May 2, 2016. The data from the Terra satellite on May 2 were not from the first pass, but they were from the first NRT remote observation. All captured data cannot be processed immediately. The processing requires spacecraft housekeeping data and predicted ephemeris. The VIIRS instrument aboard the joint NASA/NOAA Suomi National Polar-Orbiting Partnership (Suomi-NPP) satellite passed after MODIS. Hotspot locations were obtained from multiple sources including MODIS, VIRRS and AVHRR (Advanced Very-High-Resolution Radiometer) sensors. The strength of these satellite systems lies in their time-series of daily multi-spectral observations, which are used to characterize and monitor the land surface from regional to global scales. Near real-time (NRT) fire products are generated within approximately 3 h of satellite observations to meet the needs of the applicable community. There are five AVHRR satellites providing hotspots via NOAA’s Fire Identification, Mapping and Monitoring Algorithm (FIMMA) which are publicly available 3–6 h after acquisitions. We considered FIRM to represent the earliest detections to examine the potential of real-time forecasting as well.

Although each spot represents almost a 1 km gridded pixel, this does not mean that the fire covered the entire grid-cell. We consider a 25 m radius from the center of the hotspot for each ignition. We also used burned area images for 24 h fire perimeter evaluation. When vegetation burns, the burned area or fire-affected areas become characterized by deposits of charcoal and ash; fire removes vegetation or alters the vegetation structure. If soil is exposed after a fire, the brightness in Band 1 will slightly increase, but that may be offset by the presence of black carbon residue. MODIS (Terra) Land Surface Reflection (Bands 7, 2, 1 and Band 1-2-1) products are provided by NASA WORLDVIEW (https://worldview.earthdata.nasa.gov). We used the MODIS Land Surface Reflectance products available from both the Terra (MOD09) and Aqua (MYD09) satellites. The sensor resolution was up to 250 m (the imagery resolution is 250 m) and the temporal resolution was daily. As burned area products do not have a fine temporal resolution (daily images), it was not possible to evaluate the hourly location of the fire perimeter via satellite products. There was also no access to the actual fire perimeter from in situ observations (a very common observational problem). Therefore, the observations would only be sufficient for overall (daily) verification.

The only weather station close to the area of study located at Fort McMurray airport (YMM). Unfortunately, the station is far from the location of the fire, and the event date has many missing records in time. The recently released report (WMF009-2016, completed April 2017) demonstrates that the fire was reported at 16:02 on May 1. The WRF ignition algorithm uses the hotspots (the only available data at the time of simulation and the earliest available data with the current arrangement) found 5 h after the ground-based detection. This case is a prime example of the importance of collaboration in the multi-levels of provincial, national and international authorities. The M009-16 report also demonstrated that the fire was a crown fire with a speed of 10 m/min at the time. The real ignition time was not discussed in the report. However, we suspect that the ignition did not start long before detection, namely 1–2 h earlier. The daily burned area map provided by the report shows that the fire start location progressed toward the west on May 1 and then turned toward the east later on May 2. The daily changes in the burned area were driven by the wind direction changes.

4. Results

The Fort McMurray event was the costliest natural disaster in Canada and a fitting case to investigate the viability and serviceability of using a simulation to save lives and costs associated with a wildfire. However, this was also a difficult case to study due to the limited data, which is a common issue in extreme events. During this study, the only available data included the satellite images, hotspots and total daily burned area images. The WRF-Fire simulation was set up for five domains (Figure 1) for a period (lead time) of 24 h. The 24 h hindcast simulation took 6 h to complete using 144 computational processing units (CPUs). The reasonably fast performance shows the functionality of the model for real-time forecasts. The simulation started a few hours before the fire, early on May 3. The final stage of simulation at the time of pyroconvection onset is presented here. Figure 2 is an illustration of hotspots, reported fires and vertical profile locations. The vertical profile location is in the vicinity of the fire front on May 3. The model wind speed output is illustrated in Figure 3. Wind speed increased in the burned area and reached its maximum where the heat flux was at maximum, as illustrated in Figure 3b. The convergence and divergence became visible and more pronounced in the finer-resolution inner domains. Figure 3a demonstrates the 1 km resolution wind at 10 m above-ground at 00 UTC on May 3, 2016. Figure 3b simultaneously shows the 200 m resolution wind speed and direction. The comparison of the two images shows that the wind speed at the burned area was up to 3 times faster than the surrounding environmental wind speed.

Figure 4 is a 3D illustration of the vertical wind speed (updraft). The north–south view of Fort McMurray city demonstrates many up and down draft cells in the burned area. The strongest updraft is presented in red, and downdrafts are in the blue spectrum. The vertical velocities between −1 to 1 m/s were eliminated by the reduction of color transparency in this figure, allowing a better presentation of severe pyroconvection updrafts. Rising air occurs in the region of maximum vorticity, and sinking air is associated with negative vorticities. The next figures shows variational views at the location of the strongest model output for updraft.

Figure 5 is a thermo-diagram (Skew-T diagram) of model outputs at the location of 56.69${}^{\circ }$ N and 111.39${}^{\circ }$ W (a few kilometers to the north of the burned area). The thermo-diagram shows a well-established environmental convective available potential energy (CAPE) by 00 UTC on May 3. Figure 6 shows two thermo-diagrams at the same location but at a different time to show the atmosphere profile changes due to daily warming and synoptic forces. The simulation demonstrated a rapid change in environmental CAPE starting at 18 UTC on May 2. The CAPE reached 887 J/kg in less than 5 h (Figure 5 and Figure 6b). The unstable atmosphere was caused by the diurnal heating and moistening of the lower planetary boundary layer (PBL) While the nocturnal thermal inversion disappeared at around 18 UTC (Figure 6a), the dew point temperature increased at the lower PBL via synoptic moisture advection. The hourly observation from YMM (Fort McMurray airport) demonstrated the same trend of temperature and moisture increasing with slightly lower dew-point temperatures as model outputs.

The wind direction changed around 18 UTC from south westerlies to south easterlies in the burned area. The high bulk Richardson number of 93 (BRN) at 23 UTC (Figure 5) was an indicator of the small vertical wind shear and high CAPE; it was also an indicator of strong buoyant convection. The mid-level lifted index (LI, which is an index used to assess low level parcel stability) changed from positive ($1.5$) to negative ($-4.7$) during the same period of time. A negative LI indicates that the PBL is unstable with respect to the middle troposphere. An LI magnitude of $-4$ to $-7$ is classified as a large, unstable BPL. This is an environment in which convection/pyroconvection can occur. The more negative the LI, the more unstable the atmosphere will be for a rising parcel of air from the PBL. At the same time, the lifting condensation level (LCL) and level of free convection (LFC) merge at the 800 mb pressure level (2–2.5 km above the ground) with an equilibrium level (EL) about 270 mb or 11 km above ground level. Our simulation indicated that the pyro-cumulus cell onset was as late as 21 UTC on May 2, when the CAPE reached 650 J kg${}^{-1}$, and the convection inhibition (CIN, which is the energy that is needed to lift an air parcel upward adiabatically to the LCL) was 0 J kg${}^{-1}$. The dew-point temperature of the atmosphere profile above 50 kPa increased to a maximum of 10 ${}^{\circ }$C around 8 km above-ground in less than 13 h (Figure 6a,b, green line). Moistening the entire atmospheric profile increased the atmospheric buoyancy.

Figure 7 gives a demonstration of three cross-sections at the grid points [170, 186], where the highest heat flux and wind speed occurred. The figure includes three snapshots at 00 UTC on May 3, 2016 in three different vertical planes at 163, 0 and 90${}^{\circ }$ from north at the the grid point with the highest wind speed. Figure 7a is identical to a classical MCS cross-section with two downdraft areas behind and in front of the main updraft cell. The extension of the updraft 5 km above ground level shows the extension of pyro-cumulus in the back and front of the main cell. The two other cross-sections (Figure 7b,c) are the meridional and zonal views of the cell at 00 UTC on May 3, 2016. The pyroconvection activity slowed above the freezing level, but updrafts stretched to the top of the troposphere (Figure 7b,c). The vertical updraft reached the tropopause as early as 6 h after the fire ignition in the model (not shown here).

Figure 8 and Figure 9 are two close-ups of the horizontal planes at model levels 1 (second level of 60 model levels) and 10 (boundary layer). Figure 8 displays a detailed perspective of horizontal wind, temperature and humidity slightly above-ground within the center of the pyroconvection updrafts and downdrafts. Figure 9 is similar to Figure 8 at model level 10, located around 2.5–3 km above the ground. The comparison of the vertical and horizontal views indicates the developed cells, associated downdrafts, dry air, wind convergence and divergence locations. The thermal gradient aloft reached 2 ${}^{\circ }$C at 3.5 km above the ground (Figure 9b). The thermal gradient created strong convergence and divergence areas, strong surface winds, updrafts and downdrafts (Figure 8a and Figure 9a). The updraft reached a magnitude above 11 m/s at around 4 km of altitude (Figure 7c), which is well above the LFC (Figure 5). The unstable PBL with the fire wind created large cells of pyrocumulus. The downdrafts of the main cell created smaller cells in the direction of the lower-level (0–6 km above the ground) vertical mean wind. On the other hand, the heat fluxes from the fire front created a horizontal thermal gradient as large as 10 ${}^{\circ }$C in a distance of less than 200 m (Figure 8b) at the surface.

Figure 10 demonstrates the burned area on May 1 and 2. Figure 10a presents the WRF-Fire burned area outputs at 00 UTC on May 3. The overlay of the May 2, hotspots and WRF-Fire burned areas is acceptably matched. WRF-Fire outputs gave a demonstration of the rate of spread, the fire perimeter as wind direction changed and pyroconvection activities. They captured the change of wind direction and timing when the eastward fire front slowed down and the fire progressed westwards towards the river. The simulated fire front shifted almost 2 km west of the actual fire perimeter in this simulation on the west and east sides. The bias is mainly due to the ignition locations, time and biomass map resolution, which is discussed in Appendix A.

The NASA image on May 3, 2016 (Figure 11) demonstrates the large pyrocumulus cloud over the area with a distinguished rotation and overshooting area. Figure 12 shows two different views of the Horse River wildfire at early development time (23:51 UTC May 1). The updraft and downdraft can be identified at early stages. There is no report on lightning during this fire (a strong downdraft can produce more fires via lightning).

5. Conclusions

Despite the shortcomings of initialization (hotspot acquisitions) and high-resolution vegetation and biomass fuel data, we were successful in simulating the pyroconvection activity with good accuracy in a timely manner. The 24 h simulation was completed in 6 h using 144 computational processing units (CPUs) from the resources of Compute Canada. We captured the increased horizontal wind speed and updraft, moisture fluxes, strong thermal gradient, updraft and onset of pyroconvection. None of the above could be captured with an atmosphere or wildfire model. A coupled atmosphere–wildfire model is required.

The use of the vertical atmosphere profile is crucial to foresee the onset of pyroconvection, unstable atmosphere and buoyancy. Model outputs can be a suitable substitute for the atmospheric profile around a burning area where there is no sounding. The simulation demonstrated that the combination of the day-time elimination of nocturnal capping inversion, Horse River wildfire pyroconvection (increasing heat and moisture at the surface) and moisture advection increased the CAPE. The 887 J/kg total increases of CAPE in less than 5 h and the high bulk Richardson number of 93 (BRN) were indicators of the growing pyro-cumulus cell. The simulation showed the adequacy of the current WRF-Fire approach for real-time applications.

We have made an effort to overcome all compatibility, technicality and mismatch issues between different data sources and the WRF-Fire model, such as adapting the Canadian biomass fuel classification for WRF-Fire. However, there is a larger stumbling block to overcome on the road to an operational coupled model. Wildfires and associated problems are classified as ground-based issues, which are managed by provincial, territorial or state levels of governments. Unlike fire, the atmosphere is a concern for national and international organizations and management levels. The separation of authorities and their abilities inti the two key fields is an obstacle on the road to an operational coupled atmosphere–wildfire model that requires actions and collaborations from outside of scientific communities. While it sounds simple to add a fire model to a weather model (and it has been done in the past), in a complicated and well established organized NWP developing centre, no national or international weather forecasting center has crossed into the area of wildfire forecasting. As a result, natural resource organizations with limited meteorological staff or developers focus on 2D spread models, and weather prediction organizations use a dispersion model approach for wildfire emission forecasts with biases due to a lack of wildfire heat, moisture and emissions coupling. Similar to weather and climate, wildfires know no national boundaries; thus, national and international cooperation at a large scale is essential for the development of an advanced operational coupled atmosphere–wildfire prediction system equipped with data assimilation and numerical weather forecast experts. Wildfire management is done mainly by provincial authorities in Canada, and they receive national help in times of emergency at their request. However, it unclear how they can foresee the requirement help without an accurate prediction method. No province in Canada has the potential to adopt an operational coupled weather–wildfire model simply because of the lack of an operational numerical weather prediction (NWP) center with data assimilation capacity and sufficient expertise. The temporary solutions in the past, such as short-term contracts, graduate research grants and IT developers, did not result in an evolving operational solution that can be updated with advancements in observation and fire science.

Regarding the areas of improvement within the current paper, we recognize that the verification of a model is not scientifically possible at this time due to limited data and accessibility. We also realize that satellite fire detection has limitations including time lapses, false alarms and undetected fires or misrepresentations of the actual burned area and fire size. Additionally, we found that the fuel classification can be improved by using a dynamic approach with high-resolution vegetation satellite images; this will be another research topic in future. The uncertain initialization and initial conditions lead to biased results, and the bias increases by the lead time in simulation. Advancements in initialization improve the overall outcomes.