An offline coupling of fire spread models to simulate the 2021 Marshall Fire

WRF-Fire

WRF-Fire (Coen et al. 2013) couples the Advanced Research WRF (ARW) model (Skamarock and Klemp 2008; Skamarock et al. 2021), which is a widely used numerical weather prediction (NWP) model designed for both research and operational applications, with a wildland fire behaviour model based on the Coupled Atmosphere–Wildland Fire Environment Model (CAWFE) (Clark et al. 1996a, 1996b; Coen 2013). WRF-ARW is capable of multiscale atmospheric simulations ranging from mesoscale to microscale. This multiscale simulation capability together with other capabilities of WRF (e.g. detailed land surface parameterisation, radiation, microphysics schemes) can enable WRF to provide localised weather predictions with higher spatial and temporal resolution than the currently available weather data for fire simulation (Shamsaei et al. 2023).

Using fields resolved by WRF’s atmospheric grid, the fire behaviour model simulates surface fire spread with Rothermel’s rate of spread (ROS) model (Rothermel 1972) and a level-set algorithm to track the fire on the surface (Munoz-Esparza et al. 2018). After propagating the fire at each timestep, the fire behaviour model calculates fire-generated sensible and latent heat fluxes based on the combusted fuel. The calculated values are sent back to WRF’s atmospheric grid as temperature and moisture fluxes to simulate the effects of fire heat on atmospheric circulation, allowing the fire to ‘create its own weather’. WRF-Fire has been used for both prescribed burns and landscape-scale historic fires in various studies in the literature (Thomas et al. 2017; DeCastro et al. 2022; Juliano et al. 2023; Shamsaei et al. 2023). WRF-Fire is validated for a number of prescribed burns and wildland fires, especially when fuel characteristics (mass loading, moisture content, etc.), atmospheric winds and topography are quite homogeneous. When any one of these factors becomes increasingly heterogeneous, additional errors may arise in WRF-Fire’s spread prediction. It is important to note that neither canopy fire spread nor fire suppression is presently represented in the model.

SWUIFT

The SWUIFT model (Masoudvaziri et al. 2021) simulates fire spread within WUI communities by simplifying the domain of study into a raster with three different fuel categories: structure, vegetation, or non-combustible. Starting from initially known fire positions, SWUIFT incorporates thermal radiation and fire spotting mechanisms to spread the fire across the domain at regular time intervals (e.g. 5-min increments). The model assumes structure cells to be independent compartments with a height of 3 m and a timber frame construction. The structure cells can ignite by thermal radiation if they are exposed to a radiant heat flux above 14 kW m^–2, or by fire spotting if more than 24 g of landed firebrands accumulate within a 78.53 cm² circle. Once a structure cell’s fire develops, it emits thermal radiation and firebrands. However, vegetation cells ignite if they receive more than 224 g of firebrands, and they can only propagate the fire via fire spotting. The simulation process that has been described is represented in Fig. 1, and further details can be found in Masoudvaziri et al. (2021).

Fig. 1.

Simulation process followed by SWUIFT at each timestep.

SWUIFT has been used to analyse historical WUI fires such as the 2007 Witch and Guejito Fires (Masoudvaziri et al. 2021), the 2017 Tubbs Fire (Masoudvaziri et al. 2021, 2023), the 2023 Lahaina Fire (Juliano et al. 2024) and the 2018 Camp Fire (Szasdi-Bardales et al. 2024). In all the cases, the simulations provide reasonable estimations for the ROS and the expected location of damaged structures inside the analysed communities. The analysis in the first three cases focused only on fire spread within WUI, whereas the Camp Fire study proposed and evaluated a methodology for smooth fire transition between wildland and community settings where weather radar data were used to track fire propagation in the wildland. Since the analysis of these fires, SWUIFT’s code has not undergone changes in its main fire spread mechanics (i.e. ignition thresholds, structure fire development, or the attributes of the different modes of fire propagation). Nevertheless, the larger scale of recent incidents and the availability of new data and tools have resulted in code modifications that give SWUIFT the capability to process new input layers, including boundaries for domains with different simulation rules, gridded wind velocity and dynamic fire position inputs that can be independently updated at each timestep. These modifications enable SWUIFT to integrate with fire position observations or the output of wildland fire spread models.

The current study employs the methodology that was recently developed for the Camp Fire study, although in this case, WRF-Fire simulates the fire spread within the wildland, making it the first documented case in which the capabilities of fire spread models are complemented to generate a complete simulation of a wildland–WUI fire. Specific details of how the new methodology is applied in this study are provided in the Simulation setup section.

Marshall Fire

On 30 December 2021, a destructive wildfire impacted Boulder County, Colorado. The fire caused 2 deaths, destroyed 1084 homes and 7 commercial structures, and damaged 149 homes and 30 commercial structures (Boulder County 2022a). The Marshall Fire was first reported ~11:05 am local time (UTC–07:00) burning in the vicinity of Marshall Road in unincorporated Boulder County (Boulder County 2022b). The fire rapidly spread eastward through dry grass towards the town of Superior and the city of Louisville, where most of the structural damage was concentrated (Colorado Division of Fire Prevention and Control (DFPC) 2022).

A set of fire weather conditions that facilitated Marshall Fire ignition and rapid expansion were in place on the day of the incident (Fovell et al. 2022; Benjamin et al. 2023; Juliano et al. 2023). The National Weather Service (NWS) (2022) identified strong winds and dry conditions as the two factors that played the most critical role in the disaster. On the morning of the incident, wind gusts up to 45 m s^–1 (100 miles h^–1) were observed in the area. Additionally, an atypically dry fall (autumn) and early winter of 2021 resulted in the accumulation of dry vegetation, mainly grass and shrubs, that fuelled the expansion of the fire through the wildland. Early fire suppression efforts were overwhelmed by strong winds, which not only accelerated fire spread, but also generated conditions that limited emergency responders’ visibility and compromised their safety. Evacuation orders were put in place to move the residents of Louisville and Superior away from the path of approaching fire (Colorado Division of Fire Prevention and Control (DFPC) 2022; Boulder County 2022b). The fire first affected structures in the Marshall area, but later spread east and reached the commercial and residential structures on the west side of US Highway 36 (US 36) in Superior. The fire jumped over US 36 and continued to move towards Louisville, suggesting that spotting played an important role in fire propagation during the incident.

WRF-Fire setup

The WRF-Fire’s model setup used in this study is identical to the one presented in a recent paper by Juliano et al. (2023), in which the wildland spread and meteorological drivers of the 2021 Marshall Fire were investigated. The model utilises two one-way nested domains (Fig. 2) centred at (39.944°N, 105.135°W), with grid sizes of 1 km and 111.11 m for the parent and child domains, respectively, together with 45 vertical levels, which are stretched with altitude. The parent domain covers most of Colorado and southern and western parts of Wyoming and Nebraska, whereas the child domain covers a smaller area limited to Taylor Mountain, CO, on its northwest corner and southern Aurora, CO, on its southeast corner. The initial and lateral boundary conditions for these domains are obtained from the High-Resolution Rapid Refresh (HRRR) model (Benjamin et al. 2016), with the lateral boundary condition provided at an hourly interval. In addition to the atmospheric domains, a fire mesh with grid resolution of 27.78 m is used inside the child domain, which is required for WRF-Fire. The fuel data for this domain are obtained from Landscape Fire and Resource Management Planning Tools (LANDFIRE) (2020), which categorise the fuels based on Anderson’s 13 fuel types, but lack proper representation of WUI fuel categories (Juliano et al. 2023). To simulate the Marshal Fire spread, two ignitions are initiated in coordinates (39.956127°N, 105.230487°W) and (39.951209°N, 105.231441°W) at 11:08 am and 12:00 pm local time (UTC–07:00), respectively. The fire spread through the wildland is simulated in this study from 11:08 am to 9:00 pm local time on 30 December 2021. Further details on the WRF-Fire model setup can be found in Juliano et al. (2023).

Fig. 2.

Location of the nested domains used for WRF-Fire simulations.

SWUIFT setup

The spatial domain for SWUIFT simulations (Fig. 3) is defined as a rectangular area of 158 km² limited by the following coordinates: 40.004° (to the north), 39.898° (to the south), −105.105° (to the east) and −105.263° (to the west). This domain was selected to fully cover the observed fire perimeter, the city of Louisville and the town of Superior. The spatial domain is divided into a raster of 13.89 by 13.89 m cells, with cells categorised as structure, vegetation, or non-combustible. These dimensions are compatible with WRF-Fire’s child domain’s and fire mesh’s grids (i.e. 1/8 and 1/2 of their cell size, respectively). The vegetation landcover information is obtained from the same LANDFIRE fuel maps used in WRF-Fire domains (Landscape Fire and Resource Management Planning Tools (LANDFIRE) 2020), and the location of structures is obtained from Microsoft’s US Building Footprints database (Microsoft 2022). The temporal domain comprises the period between 11:10 am and 9:00 pm local time on 30 December 2021, with the first fire location input taken from WRF-Fire at 11:10 am. The simulation is run at 5-min timesteps.

Fig. 3.

Extent of the spatial domain used for SWUIFT simulations.

For the coupled simulations (community fire spread), three scenarios are analysed:

Fig. 4.

(a) Wood light-frame residential structures usually have combustible exterior materials and small defensible space around the building, and may not incorporate active fire protection. (b) Non-residential structures usually have fire-resistant exterior materials, are surrounded by large defensible spaces due to parking lots and sidewalks, and are more likely to have active fire protection (e.g. fire sprinkler systems) in place.

Post-incident damage reports provide evidence of the improved performance of non-residential structures, as the proportion of destroyed versus damaged commercial buildings (0.2:1) is significantly lower in comparison with that of residential buildings (7.3:1) (Colorado Division of Fire Prevention and Control (DFPC) 2022; Boulder County 2022a). Fig. 5 shows the location of structures removed for the third simulation scenario, and Table 1 summarises the differences between the three coupled simulation cases that are analysed in this study.

Fig. 5.

Structures removed from the spatial domain (red) and replaced with non-combustible cells for simulation scenario HRW-NFRS.

Offline coupling of models

To enable a smooth fire line transition between the wildland and communities, the ‘three-domain solution’ proposed by Szasdi-Bardales et al. (2024) is implemented. This procedure divides SWUIFT’s spatial domain into three different regions with specific rules. As shown in Fig. 6, fire spread in the ‘wildland domain’ is exclusively dictated by WRF-Fire. In the ‘transition domain’, WRF-Fire simulates the fire spread, while SWUIFT’s fire spotting module is activated to track the embers that could potentially land and cause ignitions inside the communities. Finally, fire spread inside the ‘community domain’ is exclusively simulated by SWUIFT. In this study, SWUIFT and WRF-Fire run asynchronously on different platforms. The inter-model communication occurs when the fire position output and the gridded wind velocity from WRF inform the SWUIFT simulations. These outputs are represented in a higher resolution format in SWUIFT because it operates at higher spatial and temporal resolutions than WRF-Fire (Fig. 6).

Fig. 6.

Three-domain solution applied to an offline coupling of WRF-Fire and SWUIFT (NWSC: NCAR-Wyoming Supercomputer Center; FBFM: Fire Behavior Fuel Models; HRRR: High-Resolution Rapid Refresh model; MS: Microsoft).

The definitions of the boundaries for the community and transition domains follow the Neighbourhood-Based Housing Density (NBHD) method, also proposed in Szasdi-Bardales et al. (2024). In this method, each structure (regardless of its area) is represented as a point located at its geometrical centre. Subsequently, a density raster is built by defining a squared-shaped neighbourhood around each cell and dividing the number of points inside the neighbourhood by its area. This procedure is illustrated in Fig. 7.

Fig. 7.

NBHD method: (a) buildings are represented by their footprints; (b) each footprint is simplified into a point located at its geometric centre; and (c) a density raster is generated by assigning each cell the density of its square-shaped neighbourhood.

The NBHD method is applied in the current study using a raster of 50 m cells and square-shaped neighbourhoods of 500 m per side (0.25 km²). The community domain is defined as all the areas with a minimum density threshold of 1 structure acre^–1 (247 structures km^–2), with a minimum size of 1 km². The transition domain is defined as a buffer of 1.4 km around the community domain. This domain is implemented for computational efficiency (i.e. to prevent the generation and transportation of firebrands that are not expected to land inside the community domain). The buffer is defined such that no more than 1% of the firebrands generated beyond this distance within the wildland domain could be expected to fall inside the community domain, even under the most favourable wind conditions. The same buffer size is used for all the fire scenarios. Fig. 8 shows the areas covered by the community, transition and wildland domains for the Marshall Fire case study.

Fig. 8.

Community, transition and wildland domains for the coupled simulations.

The generated community domain covers the location of the main clusters of structures while leaving out several buildings located west of Louisville and in the commercial area along US 36 in Superior. This is owing to the large size of several commercial buildings, which reduces the density in the region because the NBHD method does not account for building areas. The low density of structures outside the community domain is expected to result in fire behaviour that is better represented with a wildland fire model than an urban fire model. In the coupled set-up, any structure located outside the community domain can ignite when reached by the fire spread from WRF-Fire, and by fire spotting if located inside the transition domain. Further analysis of the proposed methodology and its influence on the simulation results is provided in the Results and discussion section, where a sensitivity analysis is performed to show how varying neighbourhood sizes and minimum density thresholds impact on the precision of the simulation results.

Timeline of fire progression

Detailed documentation of fire progression inside communities affected by wildfires can provide vital information to understand fire behaviour and can be used to validate simulation results. Although satellite observations can be used to track fire perimeters in the wildland, they lack the temporal and spatial resolutions needed for fire tracking inside communities (Lareau et al. 2022). Meanwhile, data documented by official agencies and ordinary citizens can offer an alternative for detailed tracking of the fire progression inside the affected communities.

A collection of more than 550 videos recorded during the Marshall Fire is publicly available (KUSA-TV 2022). The compilation of these videos focused on broad documentation of the events during the incident rather than on the fire spread pattern, but several of the videos show structures, vegetation, or other elements ablaze. Thus, as part of this study, all the videos were analysed, and the location and time of active fires were manually extracted.

Fig. 9 presents 193 active fire locations classified by fire type and time of observation. All the locations were visually confirmed by comparing the videos against Google Street View images. Each mark on the map represents an observed fire location rather than the location from which the video was recorded (in many cases a single video resulted in several individual fire locations). All times are taken as reported in the videos. When a fire is observed in the same location at different times, the earliest time is reported in Fig. 9. Reported times must not be assumed to be ignition times.

Fig. 9.

Documented active fire times and locations during the 2021 Marshall Fire, extracted and verified from publicly available videos (KUSA-TV 2022).

Fire suppression activities

A summary of fire suppression activities observed during the Marshall Fire is presented in Fig. 10. Twenty-one of the locations were manually extracted from the videos that were analysed to obtain active fire locations (KUSA-TV 2022), whereas the rest of the points were collected from other sources (Brennan 2022; Fischer et al. 2022; Boulder County 2022b).

Fig. 10.

Location of documented suppression activities during the 2021 Marshall Fire.

Damaged buildings

An inventory of affected residential and commercial structures is available (Boulder County 2022a). In contrast to the reported active fire locations, this source provides spatial information of all affected structures, but without indicating the times of ignition. Therefore, this inventory is appropriate only for the assessment of the final simulation results (fire line location at 9:00 pm). Fig. 11 displays the location of destroyed or damaged structures.

Fig. 11.

Reported location of damaged and destroyed structures after the Marshall Fire; map generated based on data from the Boulder County (2022a).

Wildland fire spread

The wildland fire spread results generated exclusively by the WRF-Fire simulation are presented in Fig. 12. In this case, no structure ignitions are allowed within the wildland domain as the ignition of structures is a feature that will be analysed with the coupled simulations.

Fig. 12.

Predicted fire spread according to WRF-Fire simulation only at (a) 1:00 pm, (b) 3:00 pm, (c) 5:00 pm, (d) 7:00 pm, and (e) 9:00 pm.

A comparison between Figs 9 and 12 shows that the WRF-Fire simulation captures well the initial eastward spread of the fire. By 1:00 pm (Fig. 12a), simulation results agree with the documented fire locations (Fig. 9, green and cyan symbols) that indicate that the fire has reached Superior and crossed US 36. At later times, the simulation results present different discrepancies with respect to the documented fire locations. These discrepancies will be relevant for the discussion of coupled simulation results because any errors from the wildland simulation are propagated into SWUIFT’s urban fire simulations. In the case of the wildland fire simulation performed by WRF-Fire, most of the factors contributing to modelling errors that were mentioned in the Background section do not apply for the Marshall Fire (the fire spread in primarily short grass fuels homogeneously distributed across the landscape, canopy fire spread was not important and topography was mostly uniform over the region). The fire spread was largely controlled by unusually strong and turbulent winds, including an atmospheric hydraulic jump (i.e. wind discontinuity) near the leading edge of the fire (Juliano et al. 2023). Although WRF-Fire allows for heat and moisture feedback to the atmospheric model, prior results suggest that this feedback had a negligible impact on the fire propagation and the hydraulic jump characteristics. Therefore, the WRF atmospheric model primarily controls the wind field in this particular case owing to the dominant atmospheric forcing. Despite the thorough validation of WRF over the last two decades, small-scale mesoscale and microscale variability is anticipated owing to turbulence and other physics parameterisations. Such deviations likely contributed to the simulation’s overestimation of fire spread south of the observed perimeter line (Fig. 12). Inconsistencies in the fuel map (e.g. representing road breaks as burnable fuel) may also have contributed to this overprediction. Nonetheless, the Marshall Fire case represents a scenario where the WRF-Fire model performs well when simulating wildland fire spread prior to reaching the built environment.

Among the most important observations is that the simulated fire does not cross US 36 in the undeveloped area between Louisville and Superior, whereas the documented locations show fire presence in this area at least since 3:00 pm (Fig. 9, magenta symbols). The discrepancy can be explained by the non-combustible barrier formed by the urban land cover in Superior. Additionally, the simulated fire spread towards the north and south is delayed with respect to the documented observations. The fire was observed inside Louisville by 3:00 pm (Fig. 9, magenta symbols), and reached the cluster of residential structures in Superior by 5:00 pm (Fig. 9, yellow symbols), but the simulation results reached Louisville and Superior at 7:00 pm (Fig. 12d) and 9:00 pm (Fig. 12e), respectively. A comparison of the results at 9:00 pm (Fig. 12e) with the official fire perimeter (Fig. 10) shows that the simulation overestimates the extent of fire spread in the South.

The WRF-Fire simulation stops close to the proposed community domain boundaries in both Louisville and Superior. This confirms that the community domain (Fig. 8) produced by the NBHD method provides an appropriate representation of the WUI areas, identifying important clusters of pixels that are considered non-burnable by WRF-Fire.

Coupled simulations: community fire spread

This section presents the results (Figs 13–15) from the three coupled simulation scenarios (Table 1). All the scenarios share the wildland fire spread simulation from WRF-Fire (Fig. 12) but differ in their prediction of fire spread inside the community owing to the variation in wind and structure fuel inputs.

Fig. 13.

Predicted fire spread for scenario LRW-AS at (a) 1:00 pm, (b) 3:00 pm, (c) 5:00 pm, (d) 7:00 pm, and (e) 9:00 pm.

Fig. 14.

Predicted fire spread for scenario HRW-AS at (a) 1:00 pm, (b) 3:00 pm, (c) 5:00 pm, (d) 7:00 pm, and (e) 9:00 pm.

Fig. 15.

Predicted fire spread for scenario HRW-NFRS at (a) 1:00 pm, (b) 3:00 pm, (c) 5:00 pm, (d) 7:00 pm, and (e) 9:00 pm.

The results from the coupled simulation for Scenario LRW-AS show that by 1:00 pm (Fig. 13a), the fire ignited commercial structures on both sides of US 36 inside the transition domain. Two hours later (Fig. 13b), the fire enters the community domain and makes significant advances inside Louisville. At 5:00 pm (Fig. 13c), the fire continues to spread inside Louisville, with a noticeable widening of the main fire front, reaching the north-east limit of the simulation domain (Fig. 3). By the end of the simulation at 9:00 pm (Fig. 13e), this scenario results in 6410 ignited structures in Louisville and Superior, excluding structures located outside the simulation domain that could have been reached after 5:00 pm. These results clearly deviate from the observed fire progression (Fig. 9) and reported damage (Fig. 10) in terms of the number (1270) and location of damaged and destroyed structures. This overestimation represents an error of 405%.

In Scenario HRW-AS, the low-resolution wind input for the SWUIFT simulation is replaced by a high-resolution input generated by WRF. At 1:00 pm, the simulation results (Fig. 14a) resemble those from LRW-AS, with the fire affecting only a small number of commercial structures on both sides of US 36 inside the transition domain. Additional timesteps (Fig. 14b–e) show a fire expansion that better resembles the observed fire spread (Figs 9 and 10). This can be the result of WRF’s ability to simulate a hydraulic jump, which resulted in slower fire spread eastward, but wider north–south spread. Scenario HRW-AS results in a final estimate of 2867 ignited structures, which represents an overestimation of 126% compared with the observed damage. Additionally, the fire in HRW-AS does not reach the limits of the spatial domain by the end of the simulation, implying that the actual decrease in the simulated number of ignited structures is potentially higher. It can, therefore, be concluded that the resolution of the wind input greatly affects the simulated fire spread and predicted damage. Using a WRF-generated wind field as an input introduces more realistic wind conditions into the simulation, improving SWUIFT’s performance.

Despite the improvement of the results through the introduction of a high-resolution wind input, the simulated fire spread inside the affected communities still overestimates the number of affected structures and the reach of the different fire fronts. As discussed earlier, the default version of SWUIFT does not differentiate the ignition criteria based on the level of structure hardening and requires manual adjustment. To address this limitation, fire-resistant structures are identified and replaced with non-combustible pixels in the simulation. Most of the excluded structures are non-residential buildings made from material other than wood, located in the commercial area between Superior and Louisville, around US 36, as well as in the vicinity of the Rocky Mountain Regional Airport, southeast of Superior as shown in Fig. 5.

The results from Scenario HRW-NFRS (Fig. 15) show a delayed fire entrance into Louisville, due to the limited number of combustible pixels within the area around US 36, where the wildland fire front approached the community domain in the other two scenarios. As a result, the fire does not enter Louisville until 7:00 pm (Fig. 15d) and it does so from a position further north. Although the WRF-Fire simulation in the wildland lags by ~4 h (Fig. 9, magenta symbols), a fire entrance into Louisville from the north results in a SWUIFT simulation that better resembles the observed fire progression (Fig. 9) and damage distribution (Fig. 11). The fire spread in Superior’s residential areas is similar to that of HRW-AS, which is expected as the removal of fire-resistant structures did not affect this region. The fire enters Superior’s residential areas at ~9:00 pm (Fig. 15e), a 3-h delay relative to observations (Fig. 9, orange symbols) that can be explained by the slow southward wildland fire spread from the WRF-Fire simulation. The results from HRW-NFRS estimate a total of 1110 ignited structures, which represents an underestimation with a 13% error compared with the observed damage (Boulder County 2022a).

A comparison of the three coupled simulation scenarios shows how wind input resolution and building type consideration affect the accuracy of the simulation results. The most accurate results were obtained for scenario HRW-NFRS, which used a high-resolution wind field generated by WRF and considered fire-resistant structures as non-combustible. This scenario, along with the observations from Fig. 9, suggests that the improved fire-resistant characteristics of structures located in the area south of Louisville along US 36 prevented hundreds of ignitions that might have occurred had the fire entered the residential areas earlier. This observation highlights the potential applications of model coupling for analysing large-scale WUI wildfires.

Sensitivity analysis of the thresholds defining the community domain

Different neighbourhood sizes and minimum density thresholds result in different community domain boundaries, which can affect the simulated fire spread patterns and prediction results (Szasdi-Bardales et al. 2024). In general, varying the neighbourhood size creates a trade-off between the homogeneity and precision of the generated community domains (i.e. larger neighbourhoods yield more homogeneous but less precise domains). Meanwhile, the minimum density threshold can be adjusted to control how closely borders separating the structure clusters from the vegetation areas are followed. In this section, a sensitivity analysis is conducted to evaluate how varying threshold values impact the accuracy of the simulation results. Nine simulations of the HRW-NFRS scenario are run using a combination of different neighbourhood sizes and minimum density thresholds that result in the community domain boundaries shown in Fig. 16. For all cases, the transition domain is defined as a 1.4 km buffer area around each community domain boundaries.

Fig. 16.

Contours of community domains defined with different minimum density thresholds for neighbourhood sizes of (a) 250 m, (b) 500 m, and (c) 750 m.

The fire progression over time for each of the nine simulations for the HRW-NFRS scenario is compared in Fig. 17 against the observed final fire perimeter. Table 2 presents different metrics to numerically assess the accuracy of the simulation results, including the Jaccard Index J (Eqn 1) and the Sørensen–Dice Index, SDI (Eqn 2). Both indexes range from 0 to 1, with 0 indicating total dissimilarity and 1 total similarity. For both cases, A is the area inside the observed fire perimeter and B is the ignited area predicted by the simulation.

The results of the sensitivity analysis show that the accuracy of the fire spread prediction is more sensible to variations in the minimum density threshold than to changes in the neighbourhood size. Specifically, community domains defined by a minimum density threshold of 0.5 structures acre^–1 (124 structures km^–2) are found to produce particularly inaccurate predictions in terms of the number of ignited buildings compared with the other considered thresholds. This inaccuracy can be attributed to the inclusion of large areas predominantly covered by vegetation fuel inside the community domain. Meanwhile, variations in the similarity metrics between the predicted and observed fire perimeters are not as significantly affected when different thresholds are used to define the community domain. Overall, the thresholds originally selected for this study (i.e. 0.25 km² square-shaped neighbourhood and a minimum density threshold of 247 structures km^–2) provide an accurate fire spread prediction that offers a balanced precision between the estimated number of ignited structures and the similarity of the final fire perimeter.

Fig. 17.

Temporal progression of the fire spread in the HRW-NFRS scenario for community domains defined with neighbourhood sizes and minimum density thresholds of (a) 250 m and 0.5 structures acre^–1, (b) 250 m and 1.0 structures acre^–1, (c) 250 m and 1.5 structures acre^–1, (d) 500 m and 0.5 structures acre^–1, (e) 500 m and 1.0 structures acre^–1, (f) 5000 m and 1.5 structures acre^–1, (g) 750 m and 0.5 structures acre^–1, (h) 750 m and 1.0 structures acre^–1, and (i) 750 m and 1.5 structures acre^–1. Note: 1 structures acre^–1 = 247 structures km^–2.