Influence of fuel structure on gorse fire behaviour

Site description and burning conditions

The research site is located at a latitude of −43.409, longitude of 171.568, and altitude above the sea level of 300 m, 15 km north of the Rakaia Gorge in Mid-Canterbury, New Zealand. The experimental burn layout is presented in Fig. 1, showing the six burn blocks, hereafter referred to as plots P1, P2, P3, P4, P5 and P6. The data presented here were acquired from the P2 experimental burn as part of the wider project (video available in: https://youtu.be/fdSvSVKFrMw), which was ignited on 2 March 2020 at 12:11 pm New Zealand Daylight Time (NZDT) and lasted 5 min. The plot was ignited as a wind-driven line fire, with the downwind edge lit by drip torch, with five igniters each lighting a short section to establish a completed line as quickly as possible. Lateral ignitions at each side of the plot were also used to keep a line-like fire front. Detailed information about UAV capabilities used to collect fire spread and fuel data is provided in the next section.

Fig. 1. 

Experimental burn layout showing plot sizes, fire spread overlay on Plot 2, fire spread direction and other instrumentation locations. The top-view picture of the flaming zone was taken from an Unmanned Aerial Vehicle (UAV) and overlaid on the map at the 1-min time point after the ignition.

Temperature, relative humidity, wind direction and wind speed were recorded every minute of the burn by an on-site 10-m weather station (yellow in Fig. 1). Information about the experimental conditions is provided in Table 1. P3 and P4 were instrumented with two 30 m sonic anemometer towers (shown in blue in Fig. 1) (SATI/3(K) series 3-dimensional ultrasonic anemometer, K-probe, 150 mm vertical and horizontal measurement path length, sampling frequency 20 Hz; Applied Technologies Inc., Loughborough, UK). The two towers were near centrally located in the burn blocks as the fireline flaming zone passed across. Infrared and visible videos (not used in this work) were also acquired from the side of the P1, P2, P3 and P4 burn experiments.

The dominant vegetation type was gorse (Ulex europaeus L.), ranging from 0.2 to 2.0 m in height. Gorse height and percentage cover were heterogeneous across the site, with some areas containing tall dense vegetation, and other areas being relatively open with short clumps of gorse interspersed with grassy patches. Apart from gorse, the overstorey contained a small component of matagouri (Discaria toumatou Raoul) shrub, and the understorey consisted mostly of grasses with a component of wild rose (Rosa sp.), California thistle (Cirsium arvense (L.) Scop) and Russell lupin (Lupinus polyphyllus Lindl).

The research site had a range of fuel types and size classes, with fuel loads and the resulting fire intensity and spread rates varying across the site owing to differences in vegetation height, density, coverage, proportion of dead material, and the presence of other species. Three different fuel classes were identified:

Within the area, a 200 × 200-m burn plot was established on flat ground (a dry riverbed with slight undulations), ensuring that it aligned with the prevailing up- and down-valley wind directions. In order to characterise the fuel composition in the plot, pre-burn fuel sampling was conducted in three 4 × 1-m subplots within the burn plot. Following the methods described by Pearce et al. (2010), fuel sampling consisted of both non-destructive and destructive measurements: (1) non-destructive estimation of average fuel height and percentage cover per square metre for each fuel class; and (2) destructive sampling, undertaken by cutting and collecting all aboveground biomass associated with each fuel class, and oven-drying it to obtain total biomass and fuel loading.

Information about height, percentage cover and fuel load of each class is provided in Table 2, with standing shrub fuels split into gorse and matagouri, understorey fuel split into grass and other understorey, and surface fuels split into woody debris and litter. Thermochemical properties such as specific yield and heat of pyrolysis were not measured as part of this work.

UAV and lidar acquisition

UAV capabilities were used for two different purposes: to record nadir RGB videos of the flaming zone, and to obtain a pre-burn high-density point cloud from Unmanned aerial Laser Scanning (ULS). In advance of any flight operations, an extensive set of Ground Control Points (GCPs) was established throughout the study site. Eight GCPs evenly distributed across the study area ensured a suitable level of accuracy of the ULS data. A further 36 GCPs were established around the individual research burn blocks to be used during the georectification of the acquired frames from the visible imagery and the alignment of the two data streams. These GCPs comprised purpose-built fire-proof steel plates that were pinned to the ground and painted in a pattern of four smaller squares of two high-contrast colours that made identification of the centre of the targets more accurate. Steel was used to allow the placement of some GCPs within the burn block. The eight GCPs for the ULS data were surveyed for approximately 3 min to obtain ~180 point fixes using a Trimble Geo7X differential GPS (Global Positioning System) unit (Trimble Inc., Sunnyvale, CA, USA) together with a Trimble Zephyr Model 2 external antenna for higher accuracy. An additional point was surveyed in the centre of the study area for a longer period to attain a highly accurate point fix. This was then used as the location for the base station of the Lidar (Light Detection and Ranging) to enable enhanced accuracy through post-processed kinematic (PPK) positioning.

Aerial video data of the flame zone was captured using a DJI Zenmuse XT2 dual thermal and RGB sensor with integrated 1/17 inch (1.49 mm) 12MP RGB camera (DJI Ltd., Shenzhen, China), mounted on a DJI Matrice 210 UAV (DJI Ltd., Shenzhen, China) (shown in Fig. 1). Video was captured in MP4 format at a frame rate of 30 fps (frames per second). Pre-burn ULS data were captured with a LidarUSA Snoopy V-series system (Fagerman Technologies, INC., Somerville, AL, USA) that incorporates a Riegl MiniVUX-1 UAV scanner (Riegl, Horn, Austria). This sensor was attached to a DJI Matrice 600 Pro UAV.

Flights for the visual capture were carried out using the DJI Pilot flight control application at a height of 250 m above ground level (AGL) to ensure full coverage of the burn block in a single frame. Flights for the ULS data capture were carried out using the UgCS flight control software (SPH Engineering, Riga, Latvia) at a height of 60 m AGL and a line spacing of 30 m to maximise characterisation of the vegetation. The resulting pulse density was 306 ppm². UgCS software was selected owing to its ability to create flight paths that follow terrain and to build in banking turns to the ends of flight lines, which help to minimise potential errors in the sensors’ Inertial Measurement Unit (IMU).

Lidar data were converted from native LidarUSA format to the universal LAS (LASer) file format in two processing steps. Initially, the raw trajectory data from the rover (type GNSS) of the LidarUSA system were post-processed with the PPK GPS data from the CHCX900B base station (CHC Navigation, Shanghai, China) in Inertial Explorer Express (NovAtel Inc., Calgary, AB, Canada). The post-processed trajectory data were then combined with the raw sensor data in the ScanLook Point Cloud Export (ScanLook PC) software (Fagerman Technologies INC., Somerville, AL, USA), and to apply boresight calibrations and lever arm offsets. The resulting point cloud was then output in LAS format. Processing of the raw point cloud from Lidar was then carried out using two pieces of software. First, the point cloud was tiled and had basic noise filtering applied using the LasTools software version 210,418 (Isenburg 2021). This tile output from LasTools was then imported into R statistical software version 4.0.4 (R Core Team 2020) and processed using a data processing pipeline developed using the LidR library (Roussel et al. 2018). First ground points were classified and then a digital terrain model (DTM) with a resolution of 1 m was derived from these points. The point cloud was then height-normalised using the DTM, more noise filtering was applied to remove spurious points, and finally a pit-free canopy height model (CHM) with a resolution of 0.25 m was calculated.

Fire perimeter tracking – rate of spread

The Fire Perimeter Tracking algorithm, explained and evaluated by Schumacher et al. (2022), and provided by Melnik (2021) was used to calculate the rate of spread (ROS) of the fire from the nadir RGB video. This method allows for high-resolution 2D pixel-wise calculation of the ROS (time-varying fire front) and further derived characteristics such as 2D maps of ROS, streamwise rate of spread (ROS_y), radial rate of spread (ROS_x) and spread direction (θ_ROS). In this work, the ROS is defined as speed of propagation, or as the magnitude of the vector composed of ROS_x and ROS_y, which is not always exactly aligned with the experimental plot, as shown in the next section. This choice was made to reduce the uncertainties related to the misalignments between the wind direction and the plot, as well as to account for changes of fire progression direction due to heterogeneity of fuel structure.

The method is described in detail in Schumacher et al. (2022), and is summarised below. First, the video footage was stabilised using Blender (Blender Online Community 2019), and down-sampled from 30 to 1 fps. Each frame of the stabilised 1-fps video was analysed to identify which pixels contained active flames by applying manually identified colour thresholds in the hue, saturation and value (HSV) colour space. In order to calculate the pixel-wise rate of fire spread, a 41 × 41 pixel (5.3 × 5.3 m) moving window was established across the ignition array, with every pixel of the ignition array taking its turn to serve as the centre pixel of the moving window. Within the moving window, only timestamps immediately preceding or following the timestamp of the central pixel were retained. Vectors going from the preceding timestamp pixels to the central pixel and advancing from the central pixel to the following timestamp pixels were calculated, and these vectors averaged to obtain the rate and direction of spread of the central pixel. The window then moved over and the process was repeated for the next pixel in the ignition array to serve as the central pixel. The resulting array contained the computed rate and direction of fire perimeter spread for each pixel in the burn plot.

Flame residence time

Flame Residence Time (FRT) was the second fire behaviour variable calculated in this work. It was defined as the pixel-wise duration of the observable fire from the nadir RGB video acquired with the UAV. It was calculated using a technique illustrated in Fig. 2. After the stabilisation and HSV-based flame identification previously described, two 2D maps were created: (1) an ignition map in which each pixel contained the first timestamp of active flaming in the given pixel of the burn plot, and (2) an extinction map in which each pixel contained the last timestamp of active flaming. The ignition array was subtracted from the extinction array in order to create the flame residence time array, which contained the burning duration of each pixel across the burn plot. The result was a high-resolution map containing a pixel-wise duration of the observable fire.

Fig. 2. 

Illustration of method for determining Flame Residence Time (FRT) from the pixel-wise duration of the observable fire from the nadir RGB video acquired with the UAV.

Description of fuel structure and fire spread mechanism

Fig. 3a shows an image of the fire extracted from the RGB UAV footage approximately 1 min after ignition. This figure highlights the observable flaming zone and fire front. Fire spread was characterised by rapid progressions of the fire front involving intermittent coverage and subsequent ignition of immediate preheated unburned fuel, resulting in localised progression in short periods of time. This behaviour is shown in Fig. 3a1, which highlights the change of the observable fire front during a short period of time of approximately 20 ms (white) which included instances where flames extending horizontally but then withdrew as they stood up more vertically.

Fig. 3. 

Plan view of the colour sensed with the UAV-mounted RGB videos at two timesteps: (a) approximately 1 min after ignition and (b) approximately 3 min after ignition. The expanded image (a1) shows the observable fire front for two consecutive frames, with the white line illustrating the location of the observed fire approximately 20 ms after the black line.

One of the main technical challenges related to the image and fire behaviour analysis was obscuration produced by the thick, white smoke emitted by the smouldering fuel behind the fire, as shown in Fig. 3b. It was found that smoke obstruction considerably hindered the identification of the flaming zone after ~160 m (~3 min after the ignition) in the streamwise direction (y axis), and reduced the quality of the ROS and residence time results, as discussed in the next section. In this work, the results associated with the entire duration of the fire, including the region affected by the smoke are reported. However, the analysis is limited to locations where clear identification of the flaming zone was achieved.

Fig. 4 shows the pre-burn CHM obtained for plot P2 from ULS data. The canopy height was heterogeneous throughout the burn plot, with interspersed zones of high and low vegetation. The mean canopy height was 0.75 m with a standard deviation 0.56 m. The right side of the plot was predominantly composed of low-height fuel with occasional areas of tall fuel, while the left side contained mostly tall vegetation with some interspersed short patches. The location of the fire front at different times after ignition is also shown in Fig. 4, depicted as black contour lines. Fire front position and development were found to be considerably influenced by local variability in fuel structure. Three regions (a., b. and c.) outlined in white in Fig. 4 highlight areas of low canopy height. Regions a. and b. exhibit a similar pattern, with fire fronts initially showing a convex structure (e.g. at 60 s) smoothly transitioning to a concave structure (e.g. at 180 s). This transition is highlighted with purple arrows in Fig. 4.

Fig. 4. 

In colour (blue to red): 2D map of average height of the vegetation canopy obtained from ULS acquisitions; colour values in metres. The resolution of the Canopy Height Model (CHM) map is 7.8 × 10⁻² m/pixel. In contour (black): location of fire front over time from ignition; contour values in seconds. Areas a., b. and c. (white and violet) highlight fire propagation patterns in regions of low canopy height.

Among the different complex interactions between fuel, fire and atmosphere contributing to the aforementioned behaviour, it is hypothesised that the various morphological changes of the fire during the early stage can explain some of the observed patterns. A careful analysis of the overhead videos and of the 2D ROS map showed that the fire front evolved from an initial acceleration phase with a thin flaming zone to an established phase with a constant ROS and a deeper flaming zone during the first 20–40 m (from ~60 to ~90 s). In the initial phase, the fire spread was driven by a localised heat transfer mechanism characterised by a thin flaming zone and low burning rate. Consequently, the fire front was highly driven by fuel, resulting in localised changes of fire front as a function of the canopy height. Once the fire front was established (after 90 s), with flaming zone depth reaching up to ~50 m and an observed increase in flame heights, a wider region of vegetation was affected by the emitted radiative and convective heat flux, resulting in a less localised and more ‘connected’ fire across and along the flaming zone. This connection can be linked to the quasi-steady state in which the equilibrium between the ignition and burning rate seems to take place (as further discussed in the next section). In the case of regions a. and b., it is possible that the overall inflow of wind is modified by the fuel structure, creating a ‘channel’ effect in regions of low canopy height surrounded by taller vegetation, ultimately modifying both the speed and direction of the incoming fresh air feeding the fire front. Region c. highlighted in Fig. 4 shows an ~15 m wide canopy gap starting from y ≈ 75 m. The fire front, which reached the gap ~120 s after ignition, seems to be initially disturbed by the abrupt change in fuel height, developing a convex structure localised around the surroundings of the gap. This structure seems to be conserved during the next minute over 50 m of fire progression. This behaviour could be linked to an increase in the ignition time associated with the sudden discontinuity of the fuel structure, characterised by the heat needing to reach lower vegetation levels to achieve ignition.

The influence of the fuel structure on the spatial variability of fire spread reported in this section through the description and analysis of the CHM and fire front progression maps are further discussed in the following section using ROS and FRT as fire behaviour metrics.

Influence of canopy height on ROS and residence time

Fig. 5a, b show the maps of FRT and ROS, respectively. The low canopy height regions shown in Fig. 4 are also included in Fig. 5a for reference. On one hand, FRT ranged from 10 to 90 s, with extreme values at the plot edges where drip torch ignition occurred. Regions a., b. and c. in the FRT map are also associated with low canopy height values, which may be linked to the low amount of vegetative fuel available in these regions resulting in a limited amount of energy available for combustion. This hypothesis is strengthened by the presence of regions of high canopy height associated with high FRT (e.g. bottom left, x = 50 m and y = 25 m). On the other hand, the ROS map captured the previously described rapid progressions of the fire front in the form of localised high ROS values, which could be explained by the presence of greater flame heights and longer fire activity in regions with higher canopy height, most likely associated with higher fuel load.

Fig. 5. 

Derived datasets of the burn plot where the ignition took place along the x-axis, with the x-axis showing plot width in metres, the y-axis showing the plot length in metres, and the value corresponding to: (a) the flame residence time derived from the Unmanned Aerial Vehicle (UAV) video footage, and (b) rate of fire spread also derived from the footage. The resolution of both maps is 7.8 × 10⁻² m/pixel.

Fig. 6b shows the evolution of ROS through the streamwise direction for high canopy height (top) and low canopy height (bottom). The ROS and FRT in Fig. 6a were averaged over a flame front width of Δx = 20 m, with ±1 s.d. shown in grey. The plots are accompanied by the corresponding CHM values (blue) averaged over the same width. ROS resulting from high canopy height was found to be low during the first ~20 m in the streamwise direction, where the fire front was mainly accelerating with a progressively bigger flaming zone attached to the ignition line. After ~20 m, the fire was more established, with a quasi-steady value of ROS ~1.2 m/s and a larger flaming zone, even though the averaged canopy height from the CHM considerably decreased from ~1.6 to 1.0 m. For lower canopy heights, however, the fire seems to undergo a longer period of acceleration until ~40 m, where a maximum average ROS of ~1.0 m/s is reached. Those differences suggest that canopy height could play a relevant role on ROS during the acceleration period for the experimental conditions studied in this work.

Fig. 6. 

(a) Flame residence time (FRT) and (b) rate of spread (ROS) compared with canopy height for high (top) and low (bottom) vegetation heights. The y axis represents the streamwise direction from the ignition point (y = 0 m). The black lines are the FRT and ROS averaged over a width of Δx = 20 m. The grey ribbon shows 1 s.d. for the FRT or ROS values, and the blue lines are the average canopy height.

FRT, shown in Fig. 6a, displays a different behaviour and seems to be considerably more dependent on canopy height for both studied cases. This dependency is mostly perceivable over the first half of the plot. For tall vegetation (top plot), FRT progressively decreases by approximately 30% from y = 20 to 80 m, while the CHM decreases ~25% in the same range. After y = 80 m, FRT and CHM are overall both relatively constant except for some localised changes.

The same trend can be seen in the second studied case for the region of short vegetation (Fig. 6a, bottom plot), where changes in FRT values seem to be driven by variation in vegetation height during the development of the fire. This correlation, further explored in the next section, is associated with the strong linkage between the canopy height and fuel load (Pearce et al. 2010), which is supported by results from destructive sampling measurements carried out before the prescribed experimental burns (not shown in this work). Areas with high fuel load contain a greater amount of biomass available to participate in combustion, leading to longer combustion duration and FRT. However, it is particularly challenging to assess the extent to which wind, moisture content, fuel morphology and other variables simultaneously influence the ROS and FRT. For instance, wind speed may play a relevant role in increasing combustion efficiency by providing oxygen to feed the reaction, while fuel condition (dead-to-live ratio) and particle size modify the rate at which the reaction takes place. Fuel type, which captures these differences in the properties and structure associated with different vegetation types, may also play a significant role in these low canopy height areas owing to a change from shrubs to grass as the dominant fuel carrying the fire. Grass fuels are known to produce faster spread rates, especially under high wind speeds and when dead or cured, owing to the predominance of fine fuels, high proportion of dead fuel and lower moisture contents (Cruz et al. 2022). The next section provides further insight into the complex coupling between FRT, ROS and CHM addressed in this work by exploring mathematical correlations and statistical properties describing this interaction. This approach does not aim to provide a definitive explanation or description on how this interaction works, but rather intends to bring quantitative clarity on the first-order nature of this relationship.

Empirical correlations and statistical analysis

The ROS and FRT maps shown in Fig. 5 are each composed of more than 6 million data points owing to the high resolution of the acquired video and subsequent methodology applied to derive them. The datasets were initially filtered to improve readability of the results by averaging the ROS and FRT, as shown in Fig. 7. A detailed analysis of the non-filtered dataset is shown in Figs 8 and 9.

Fig. 7. 

Scatter plot of Flame Residence Time (FRT) as a function of the Canopy Height Model (CHM) and rate of spread (ROS) (in colour). Results correspond to data located from x = 30 to 170 m to reduce edge effects, and from y = 30 to 90 m to ensure low smoke obscuration and a steady-state fire. The results were filtered using a 20 × 20 m² non-overlapping square kernel on the three maps (CHM, FRT and ROS).

Fig. 8. 

2D probability density plots of rate of spread (ROS) as a function of the Canopy Height Model (CHM) for six different locations. Plots (a–f) are shown as 2D kernel density estimations and histograms of a subsample of datapoints. The size of the regions of interest (black squares) is 20 × 20 m².

Fig. 9. 

2D probability density plots of Flame Residence Time (FRT) as a function of the Canopy Height Model (CHM) for six different locations. Plots (a–f) are shown as 2D kernel density estimations and histograms of a subsample of datapoints. The size of the regions of interest (black squares) is 20 × 20 m².

Fig. 7 shows a scatter plot of FRT and ROS (in colour) as a function of CHM. The plot was generated by applying a median filter using a basic 20 × 20 m² non-overlapping square kernel on the three maps, accounting for 10% of the experimental plot, which was found to represent a suitable quality/size trade-off. Results correspond to data located from x = 30 to 170 m to reduce edge effects, and from y = 30 to 90 m to ensure low smoke obscuration and a steady-state fire.

A linear regression analysis was performed to explore the relationship between the average FRT and the average CHM, which yielded the following equation:

where CHM is in metres and FRT is in seconds. This equation is valid in the studied range of applicability, and yields a reasonably high regression coefficient of determination of R² = 0.74 as expected from results shown in the previous section. Most of the dispersed points are associated with high average CHM values (>1.2 m), potentially due to dissociation between fuel load and canopy height, as well as due to uncertainties linked to smoke obscuration and the calculation method. The linear regression between the average ROS and CHM yielded the following relationship:

where ROS is in metres per second and CHM in metres. This equation is valid in the studied range of applicability provided in Fig. 7. The resulting coefficient of R² = 0.26 is low, showing that unlike FRT, CHM alone is not a primary variable dictating the speed of progression of the fire for the studied case. Furthermore, another potential reason explaining the observed disparity involves a high dependency of ROS on the fuel structure (e.g. local changes in CHM). This hypothesis is further explored in Fig. 8, which shows a scatterplot of the non-filtered ROS versus CHM for different 20 × 20 m² regions of interest. Owing to the large number of datapoints (64 000 points per region of interest), the plots are shown as 2D kernel density estimations and histograms of a subsample of datapoints randomly selected (1:100 or 640 datapoints). This was necessary to facilitate readability and interpretation of the results.

In general, regions with lower canopy height (average canopy height <0.5 m – plots d and e in Fig. 8) are associated with low ROS values, presenting narrow distributions centred around 1.5 m/s. Regions with higher CHM (average canopy height >1.0 m – plots a and c in Fig. 8) present wider distributions of ROS values, reaching both high magnitudes greater than 3 m/s and values close to quiescent conditions. This behaviour is associated with quick progressions or fire flickering, as previously described, involving greater flame heights and longer fire activity. Regions covering low CHM vegetation (averaged canopy height lower than 0.5 – plots d and e in Fig. 8) showed narrow distributions, suggesting that the fire progression was considerably more ‘stable’. Regions with medium CHM values (averaged canopy height between 0.5 and 1.0 m – plots b and f in Fig. 8) were found to be in a transition regime. On one hand, Figure 8f involves a low ROS region surrounded by low CHM vegetation, associated with thin flaming zones and low FRT, whereas Fig. 8b covers high ROS channel-like vegetation structure, surrounded by high CHM vegetation linked to high FRT, and associated with a considerably larger flaming zone. It is likely that these differences in fuel structure played an important role in enhancing or hindering heat transfer mechanisms associated with the rate of progression of the fire. In the case of Fig. 8b, the favourable characteristics of the surrounding vegetation can be associated with higher flames and larger flame depth, which could strengthen the radiative heat transfer required for ignition of the unburnt fuel. Additionally, it is possible that the channel-like structure enhances convective heat transfer by increasing the flow of hot gases and the surface area of exposed fuel.

Fig. 9 shows histograms and scatterplots for FRT at the locations discussed above. The distribution of FRT for all the plots is considerably narrower than in the case of ROS, showing less dispersion and in most cases a higher correlation between the variables of interest. Similarly to the ROS cases, low CHM regions are linked to low FRT values, with distributions centred between FRTs of 20 and 30 s, and high CHM regions with distributions centred between 40 and 60 s. FRT values shown in Fig. 9a are by far the highest, with FRT values ranging between 31 and 96 s. Interestingly, the results presented in Fig. 9c show a bi-modal distribution of FRT centred on 38 and 52 s, which can be associated with high heterogeneity of vegetation heights in the region of interest. In summary, results reported in Fig. 9 are in accordance with previous findings showing a clear dependency on canopy height.