Observations of wildfire spread dynamics in southern Australian grasslands

Fire behaviour modelling and error statistics

Fire danger was calculated through the Grassland Fire Danger Index (McArthur 1966) using the Mk 3 equation of Noble et al. (1980). The Cheney et al. (1998) grass fire rate of spread models for natural and grazed grasslands were evaluated against the average observed rates of fire spread. These models predictions relied on estimated moisture content of dead grass fuels using Cheney et al. (1989) parameterisation of McArthur (1960) fuel moisture tables. Fire spread model fit was evaluated through the calculation of the mean absolute error (MAE), mean bias error (MBE) and mean absolute percentage error (MAPE) (Willmott 1982). Residuals were calculated as observed minus predicted, with positive and negative bias indicating under- and over-prediction, respectively. All statistical analyses were conducted with R software (R Core Team 2024). Fireline intensity, as per Byram (1959), was estimated using the measured fine fuel load and a low heat content of 18,600 kJ kg⁻¹.

Dataset characteristics

A dataset of six wildfires burning through grassland fuel types was assembled (Table 1). Two other available wildfires were not included as they did not meet data quality requirements. The wildfires in the dataset ranged in size between 31 and 502 ha (average 187 ha). A total of 75 fire spread segments were identified. After removing fire spread segments with uncertain start or end times, or involving wind direction changes greater than 45°, a total of 68 segments were used in the analysis. Fire spread segment period and distance averaged 200 s (3.3 min) and 261 m, respectively. The longest segment run was 1067 m. Two of the wildfires, the Glenloth East (Figs 1–3) and the Tahara, each had two independent or separate fire runs (e.g. Fig. 3, Table 1). Average wind speeds and estimated dead fuel moisture content varied between 11 and 32 km h⁻¹ and 4.7 and 9.1% oven-dry weight, respectively (Table 2). Grasslands were observed to be 100% cured, although the curing level could have been lower in some localised gullies and creek lines.

Fig. 3.

Map of the Glenloth Rd fire progression. Figure highlights the mosaic of different grass and crop conditions in the path of a wildfire typical of the southern Australia grassland environment. This fire had two independent runs, with the first run (cool colours) being stopped at a road after spreading for 1.3 km; the second run (warm colours) originated at the rear northern flank of the first run approximately 26 min after the first run stopped, and spread approximately 3.7 km before being largely suppressed at a major road. These two runs were considered independent for fire behaviour analysis. Background image is from 21 December 2021.

The Grassland Fire Danger Index (McArthur 1966) varied between 12 (Steele Rd fire) and 23 (Hillside fire), values within the High (8 < GFDI < 20) and Very High (20.5 < GFDI < 50) rating classes (Table 3). Segment rate of fire spread varied between 0.0 and 21.7 km h⁻¹. Spread rates of zero refer to periods of time when the fire was held up for several minutes, at times up to 10 min (Fig. 4), by some sort of a barrier. Such barriers were eventually breached, allowing the fire to continue to propagate (Fig. 3). The average wildfire rate of spread for overall fire runs, varied between 3.7 and 9.7 km h⁻¹ (Table 2). All observations were conducted within the first hour of propagation, although we could not identify any fire spread trend that would suggest a fire was still accelerating from its initial point source ignition (Fig. A1). All overall fire runs considered were assumed to be spreading in equilibrium with the environment having previously completed the fire acceleration phase after ignition.

Fig. 4.

Variability in observed rate of fire spread in the Glenloth Road Fire. Red horizontal lines represent the average segment rate of fire spread; grey vertical lines represent barriers such as fire breaks. Fire was reported at 10:30, and the first fire behaviour observation was conducted at 10:37 hours.

Fire spread variability

The variability of the headfire rate of spread for the six wildfires analysed is illustrated in Fig. 5 (also Fig. A1). The fastest segment fire spread rate, 21.7 km h⁻¹ on the Glenloth East fire, was 4.6-times greater than the average for the overall fire run. The ratio between the maximum and the mean rate of fire spread averaged 2.6 for the eight fire runs; i.e. on average the maximum rate of spread was approximately 2.6-times the overall fire run average. This ratio varied between 1.9 and the above mentioned 4.6. The coefficient of variation of rate of fire spread averaged 88%, varying between 39 and 173% (Table 2).

Fig. 5.

Variability in observed rate of fire spread in the eight wildfires runs used in the analyses. Boxes define the upper and lower quartiles, with the thin line within representing the median. The open circle is the fire run average rate of spread. The outer horizontal lines outside the box reflect maximum and minimum values excluding outliers (black circles), defined as more than 3/2-times the upper quartile.

Rate of fire spread was dynamic or non-steady in all fires (Fig. A1). Very slow rates of spread were associated with periods where the headfire was held up by, but eventually crossing, a barrier. We could not identify a clear acceleration trend in the segments downwind of a barrier once the fire breached it. The rate of spread of the fire segment after being held and overcoming a barrier was often quite high (e.g. Fig. A1a, e, f). We were only able to observe a cyclic trend in the longer duration runs (Fig. A1b, d, e, f). Smaller runs (Fig. A1a, c, e) did not have the duration to establish time series where a cycle was observable.

Effect of barriers on grassland fire propagation

Of the 68 fire spread segments used in this analysis, a total 41 segments had fire crossing, or breaching, barriers such as roads or other types of breaks. Barrier characteristics varied from firebreaks and farm lanes, typically 3-m wide, to gullies (often with non-fully cured grasses) about 20-m in width (Table 3). Firebreaks and roads ranged from poorly maintained (e.g. some fuel might be present) to well-maintained, with surrounding vegetation managed (e.g. cut or well grazed grass).

Under the burning conditions associated with the dataset, no road, firebreak or gully stopped a wildfire per se. We observed that most poorly maintained barriers with widths of up to 3.0 m did not affect the propagation of a wildfire if it was spreading vigorously. Wider barriers, such as roads, tended to slow the overall wildfire propagation, and at times facilitated its containment.

The effect of a barrier on fire propagation was dependent on the speed and intensity of the fire at the time it reached the barrier, the width and condition of the barrier, the grass fuel condition surrounding the barrier (e.g. eaten-out vs natural or unharvested crop), and the existence of trees or shelterbelt on the windward side. Collected data did not allow for a quantitative analysis of the effect of barriers on fire propagation, although the data were in agreement with Wilson (1988) firebreak breach model.

Evaluation of Cheney et al. (1998) grass fire spread models

The grass fire spread models of Cheney et al. (1998) were evaluated using the overall fire runs (Tables 1 and 2). With the exception of one shorter fire run (i.e. MurraWarra fire, 1.8 km in 12 min), all other fire runs lasted approximately 30 min or more. Fire runs incorporated all the fire segments within its spread footprint, also encompassing periods of slow propagation due to the effect of barriers as envisioned by Cheney et al. (1998). Fire spread predictions used the most appropriate fire spread model for the prevailing grass fuel condition: cut and grazed model for mid-summer pastures (Cheney and Sullivan 2008) and harvested crops (Cruz et al. 2020); and natural grass model for unharvested crops (Cruz et al. 2020). The models predicted the dataset with an MAE of 1.72 km h⁻¹, and a MAPE of 26%. Mean bias was 1.16 km h⁻¹. Four of the eight fire runs evaluated were within the ±35% error band (Cruz and Alexander 2013). The fastest fires were under-predicted (Fig. 6a, b). Interestingly, these fires had moderate wind speeds (i.e. between 13 and 21 km h⁻¹) and very low estimated fine dead fuel moisture contents, between 4.5 and 5.2% (Fig. 6c, d). No other available variables, such as run duration (Table 1) or wind gustiness (Table 2), explained the under-prediction of the four fastest fires.

Fig. 6.

Plots of (a) observed rate of fire spread vs predictions of the Cheney et al. (1998) fire spread models, with bubble size proportional to the fire run duration (range, 12–37 min), (b) residuals vs observed rate of fire spread, (c) residuals vs 10-m open wind speed, and (d) residuals vs predicted dead fuel moisture content. The dashed lines around the line of perfect agreement in (a) indicate the ±35% error interval.

Fire spread variability

The fire variability data analysed had length scales between the observations in experimental fire behaviour studies, e.g. tens of meters and less than 2 min (e.g. Burrows et al. 1991; Taylor et al. 2004; Butler et al. 2016), and wildfire data, e.g. longer than a kilometre and an hour (Rothermel 1991; Viegas et al. 2022). All wildfire runs showed an unsteady dynamic in rate of fire spread, with the longer runs showing a cyclic variation in spread rate (Fig. A1). It is unclear if the main determinant of the amplitude and frequency of the observed cycles was due to changes in wind speed, fuel condition, suppression, or fire-atmosphere interactions; however, it was observed that in a number of occasions fires would temporarily halt while reaching barriers. The dataset does not contain the detail necessary to investigate the drivers of the cyclic fire dynamics, mainly because the wind measurements were conducted tens of kilometres from the active headfire region.

The data collected give insights into the variability in rate of fire spread expected to occur in wildfires burning in similar conditions in southern Australia, with spread rate having an average coefficient of variation of 88%. Probably of more relevance for suppression operations is the finding that on average, the peak rate of fire spread can be almost three times higher than the observed or predicted fire run average value. This value is higher than the average of 1.7 presented by Rothermel (1991) for a series of wildfires burning in conifer forests of the northern Rocky Mountains, USA; and 1.8-times observed in experimental crown fires carried out in the boreal conifer forest in the north-west Territories of Canada (Taylor et al. 2004). Sudden changes in fire behaviour have the potential to disrupt effective suppression operations (Plucinski 2019) and place firefighters in danger.

The high maximum:mean rate of fire spread ratio observed in our dataset suggests a higher amplitude in rate of fire spread cycles in grassland fuels. This is likely because the prevailing open environment, very fine dead fuels, and relatively smaller energy release rates make the flame front more reactive to changes in wind, which has implications for fireline safety and risk analysis. It is important to emphasise the relatively small size of the dataset used in this analysis limits any more definitive conclusions. The effect of barriers to spread in grassfires, causing fires to momentarily stop at a location, will also reduce the average rate of spread of a fire run, which in turn will result in a higher max/mean rate of fire spread ratio. Analysis of a larger dataset, such as that of Cadril et al. (2023), will be necessary in order to more comprehensively explore the variable behaviour of wildfires.

Effect of barriers to fire spread

Data characteristics, namely the limited size of the database, did not allow for a robust quantitative analysis of the effect of the barriers on fire propagation. Despite this fact, the observations in the dataset did allow for a number of qualitative inferences. Landscape fuel discontinuities, in the case of southern Australian grasslands being mostly barriers such as roads, have a visible effect on overall fire propagation and suppression effectiveness (McArthur et al. 1982). Wider barriers, such as established roads 5–10 m wide, tended to significantly slow the overall wildfire propagation, and at times facilitated containment. When a wildfire reached such a barrier, the headfire would stop momentarily, with discrete ignition points observed to occur downwind of the barrier. It was observed that such ignitions could occur immediately or take minutes to materialise; up to approximately 10 min in some cases. This process was often repeated several times in the longer runs (e.g. Glenloth East fire, Fig. 4). This momentary halting of overall fire propagation and associated reduction in fireline intensity often allowed for successful containment of the fire at that location. However, we could not identify situations where the barriers, at times consisting of 15-m wide tracks, stopped a fire solely by itself. From the observations, it was evident that small wind shifts on the flanks or back of the fire allowed a new headfire to develop and breach the barrier; in some cases, this occurred on multiple occasions before the barrier was finally breached. It is uncertain as to whether these breaches across barriers were due to wind gusts increasing the flame length and thereby enabling direct flame contact across the barrier, short distance ember transport, or a combination of both.

Small barriers, such as 3-m wide roads, showed negligible effect on slowing down the headfire region, with flame fronts often progressing unimpeded across these barriers without any observable variations in rate of fire spread.

Successful containment with ground resources supported by aerial firefighting often occurred after the flame front reached a barrier and was momentarily held up. This highlights the fact that these areas had the highest chance of successful fire suppression, with this occurring even under conditions conductive to high intensity fire propagation.

The observations discussed here have implications for simulating landscape fire propagation in the context of how the existence of barriers to fire spread and their impacts are considered. If a simulator uses an empirical fire spread model such as Cheney et al. (1998), a model that implicitly includes the effects of barriers in momentarily slowing down wildfire propagation (see next section), then the simulator should not explicitly include the effect of the barriers in fire propagation. Doing so can result in doubling the barrier effect in slowing down fire propagation, leading to an under-prediction in rate of fire spread. Considering practical applications, this is likely only relevant for low to moderate burning conditions. Under heightened fire danger conditions, the resultant higher fireline intensities will be above most thresholds for barrier breaching (Wilson 1988) and the simulations should overcome barriers without slowing the fire down. Testing of spatially explicit wildfire propagation simulators with the present wildfire dataset will improve our understanding of how wildfire growth simulators should incorporate the effect of barriers on fire propagation.

Evaluation of current models

The evaluation of the Cheney et al. (1998) fire spread models yielded a mean absolute percent error of 26%, a relatively low value when compared with other results obtained from the evaluation of grass fire spread models against wildfire data, such as a value close to 80% for the same models in Kilinc et al. (2012; also Cruz et al. 2018a) or just below 40% in Cardil et al. (2023). Previous evaluation of the models against experimental data in pastures and croplands (Cruz et al. 2018b, 2020), with wind, fuel structure and fuel moisture content measured, produced percent errors varying between 22 and 35%, an error range similar to what was found in our analysis. See Sneeuwjagt and Frandsen (1977), Rothermel and Rinehart (1983), van Wilgen and Wills (1988) and Gould (1991) for further studies evaluating fire spread models against experimental and prescribed fire data in grassland fuels.

The evaluation results showed an under-prediction bias, with the four fastest wildfires being under-predicted. This is likely the result of a skewed dataset rather than a model limitation. The collection of fire spread data in wildfires spreading under moderate to high fire danger rating conditions as done in the current study is subject to a bias. Generally data from experimental fire observations or unsuppressed wildfires will be symmetrically distributed around the mean (expected prediction from model; e.g. fig. 4 in Cheney et al. (1998)). But in the case of the present study, the wildfires that spread faster than the expectation (above the mean expectation for the conditions) are more likely to escape initial control, propagate across the landscape and in turn be documented. Conversely, the fires spreading slower than the mean expectation are more likely to be contained while small, and thus not be subjected to documentation as a free spreading wildfire. As such, a dataset such as that assembled for this study will be biased to the fires spreading faster than the mean expectation. The skewed nature of the dataset due to the absence in the analysis of some fires spreading slower than expected yielded an under-prediction bias.