The effect of fuel bed height in grass fire spread: addressing the findings and recommendations of Moinuddin et al. (2018)

Additional keywords: crop fuels, fire behaviour experiments, fire mitigation, grass fuels, headfire.

Moinuddin et al. (2018), hereafter MSM18, used simulations from a physics-based model (Mell et al. 2007) to quantify the effect of fuel height on the rate of spread of grass fires. The simulations, based on fixed environmental conditions and an artificial fuel bed resembling grass with variable height (bulk density held constant; fuel load varied proportionally with fuel height), yielded two disparate trends in fire characteristics. The authors interpreted the results as the signal of the existence of two distinct flame spread regimes, termed ‘boundary layer mode’ and ‘plume mode’, and suggested a Froude number (Fr, as calculated by Apte et al. 1991) of 0.5 as the threshold separating these spread modes. From the simulation-derived effect of grass fuel height on rate of fire spread for the so-called ‘plume mode’ (Fr <0.5), they proposed a parametric relationship between these two variables. The relationship suggested a strong and inverse effect of grass fuel height on rate of fire spread (i.e. a reduction in fuel height will result in a substantial increase in rate of fire spread). MSM18 noted that ‘…these results shed light on dependence on grass height and it is useful for fire agencies to have a correlation between RoS of the fire and grass height to assist in strategic decision-making’ (p. 811) and further suggested the broader use of physics-based model simulations to develop parametric models in support of fire management decision-making.

Although the MSM18 grass height–rate-of-spread relationship might be considered novel, it is counter-intuitive. After all, a common fire mitigation measure is to mow or cut grass to reduce the spread and intensity potential of a wildfire (e.g. NSW Rural Fire Service (RFS) 2010; Clements et al. 2019; Country Fire Authority (CFA) 2019). As they stand, the MSM18 results would necessitate a rethink of how grass fire mitigation measures should be conducted; instead of decreasing grass fuel height as a fuel hazard reduction treatment, grasses should be left in their natural state and their growth even encouraged.

In our view, it is important to comment on MSM18’s results and conclusions, namely due to their potential applied implications. At a time when computer-based modelling is suggested as a way to advance our understanding of fire behaviour and produce applied outcomes (Hoffman et al. 2018), often due to the difficulties and costs of carrying out field-based research (Alexander and Quintilio 1990), it is critical to identify incongruencies in modelling results before they become established and widely accepted (also apposite to results from field experiments and their analysis).

The objective of this comment is to verify MSM18’s simulation results through the analysis of robust empirical evidence of the effect of grass fuel bed height on the forward rate of fire spread as measured in field-scale experiments.

To analyse the effect of grass height in fire spread rate, we used two datasets (Cheney et al. 1993; Cruz et al. 2020) of free-spreading experimental fires where an explicit manipulation of grass fuel height has been conducted, mimicking the simulated treatment of fuels in MSM18. We used the Eriacchne spp. (locally known as kerosene grass) subset (71 experimental fires) of the CSIRO Annaburro Station study in the Northern Territory, Australia, as described by Cheney et al. (1993). This subset contained three grass states: E1 – natural undisturbed grass (control), and the result of two fuel manipulation treatments; E2 – grasses cut to 50% of their natural height and clippings left on site; and E3 – grasses cut to 50% of their natural height and the clippings removed.

We also used data from Cruz et al. (2020) fire spread experiments in a wheat crop (Triticum spp.) in Victoria, Australia, for three different fuel conditions: Harvested (control); Unharvested (treatment 1); and harvested and baled (treatment 2, hereafter termed Baled). Fires in this dataset were conducted as simultaneous paired experiments (control v. treatment) to observe the direct effect of differences in fuel structure on fire behaviour.

Fuels in both the Cheney et al. (1993) and Cruz et al. (2020) datasets were destructively sampled and reported on a per experimental fire basis. Readers wanting detailed information on the methods used for weather and fire behaviour measurements should consult the original publications.

The Apte et al. (1991) Froude number (Fr) was calculated for each fire in the Cheney et al. (1993) and Cruz et al. (2020) datasets utilising the methodology of MSM18.

For the Cheney et al. (1993) dataset, we analysed the relationship between fuel height and rate of fire spread through correlation and linear regression analysis. A Tukey’s post hoc test was used to conduct multiple comparisons of the three grass condition characteristics and associated rate of fire spread. For the Cruz et al. (2020) dataset, we conducted direct comparisons between the relevant variables through paired t-tests. Given the reduced sample size of each population, the Shapiro–Wilk test of normality was used to determine if the variables were normally distributed. Grassland Fire Danger Index (GFDI), a surrogate of fire spread potential incorporating the effect of wind speed and fuel dryness, was calculated using the Mk 3 equation given by Noble et al. (1980). All statistical analysis was conducted using the software R (R Core Team 2018).

Table 1 summarises the average fuel characteristics, rate of fire spread and Fr for each of the grass conditions (see also Fig. S1 available as Supplementary Material to this paper). The changes in fuel height are within the range of the simulated treatments in MSM18, with the E3 treatment most closely replicating, in the real world, the simulated treatment (i.e. a reduction in fuel height with the removal of cut fuel particles leaving fuel bed bulk density mostly constant). Fr¹ in the dataset varied between 0.11 and 0.47, with lower values associated with the experiments carried out in the undisturbed grasses (E1).

Table 1.  Average and standard deviation (in parentheses) of measured environmental variables and rate of fire spread for the three Eriachne spp. grass height conditions in Cheney et al. (1993) and the three Triticum spp. crop conditions in Cruz et al. (2020)
GFDI, Grassland Fire Danger Index; Fr, Froude number

Average grass fuel height varied between 0.31 m in the control (E1) to 0.13 and 0.11 for E2 and E3. A Tukey multiple comparison test revealed significant differences in fuel height between E1 and E2 or E3 (Table 2). No significant differences were found for dead fuel moisture or GFDI. The control had the lowest 10-m open wind of the three subsets, with the difference between E1 and E2 being significant (P < 0.05; Table 2), but not significant between E1 and E3. Despite the lower average wind speed, E1 had the highest average rate of fire spread (Table 1), but the differences between the control and treatments were not significant (Table 2).

Table 2.  Multi-comparison test (Tukey’s honestly significant difference) for changes in environmental conditions and rate of fire spread for the three grass height conditions (E1, E2 and E3) in Cheney et al. (1993)
GFDI, Grassland Fire Danger Index

Linear regression analysis of rate of fire spread with wind speed, dead fuel moisture content and height reduction treatment as a categorical variable resulted in an adjusted R² of 0.60 and indicated a reduction in rate of fire spread with treatment (Table 3), i.e. a reduction in fuel height resulted in a reduction in rate of fire spread. The coefficients for E2 and E3 suggest a bulk reduction in rate of fire spread of −0.7 and −1.0 km h⁻¹, respectively, over the control. The magnitude of this reduction is notable given the mean and range in the rate of fire spread for the dataset (Fig. S1). Notably, the effect of E2 was quantified as being weakly significant (P = 0.06), whereas the effect of E3 was quantified as highly significant (P = 0.004).

Table 3.  Coefficients for linear regression analysis for Eriachne sp. data in Cheney et al. (1993)

Paired burns (i.e. control v. treatment) allowed direct contrast of the effect of fuel height on rate of fire spread for two fuel configurations, i.e. Harvested v. Unharvested and Harvested v. Baled. Table 1 provides a summary of the range in fuel, weather, Fr and rate of spread data per crop condition (see also Fig. S2). Mean fuel heights varied between 0.73 m in the Unharvested condition, 0.29 m in the Harvested and 0.09 m in the Baled condition. Fuel load for standing fuels varied as: Unharvested: 0.53 kg m⁻²; Harvested: 0.21 kg m⁻²; and Baled: 0.13 kg m⁻². Harvesting operations in the Harvested and Baled treatments led to the formation of a matted fuel layer on the ground. The structure of this layer did not differ between the Harvested and Baled conditions, with an average matted fuel load of 0.23 kg m⁻². No matted or ground fuels were present in the Unharvested condition. Importantly, in the experimental fires the ignition interface was observed to spread in the standing fuels, with the charring and consumption of matted fuels occurring a few metres behind the leading edge of the flame front (Cruz et al. 2019). As such, despite the presence of matted fuels, the dominance of the standing fuels driving fire propagation approaches the artificial condition used by MSM18 of only standing fuels with constant bulk density. The matted fuels, when present, contributed to the overall energy release but not to the net horizontal heat flux as per Thomas et al. (1964) and thus rate of spread. All Unharvested and Harvested experiments were characterised by Fr <0.5. Fr averaged 0.5 in the Baled condition, with four fires characterised by Fr being between 0.5 and 0.58. Overall, 92% of the fires in this dataset had Fr <0.5.

Figure S2 shows the marked differences in fuel structure between the three configurations, similarities in environmental burning conditions (10-m open wind speed, fuel moisture and GFDI), and the notable differences in observed rate of fire spread. Figure 1 provides visual contrast between fires burning in the three fuel types under similar fire danger conditions. Overall, the Unharvested condition showed the highest rate of fire spread, with an average rate of spread of 7.03 km h⁻¹ and a maximum value of 10.15 km h⁻¹. In contrast, the Baled condition was characterised by the lowest rates of fire spread, with an average value of 3.26 km h⁻¹ within a range of 2.36 and 4.58 km h⁻¹.

Fig. 1.  Contrasting fire behaviour between three experimental fires in wheat crops with distinct fuel arrangement as the flame front reaches the end of the plot. Reference markers are 1.6-m tall. Top: experimental fire WU4 (Unharvested condition), 50 s after start of ignition, R (rate of fire spread) = 5.26 km h⁻¹. Middle: experimental fire WH11 (Harvested condition), 54 s after start of ignition, R = 4.12 km h⁻¹. Bottom: experimental fire WHB8 (Baled condition), 81 s after start of ignition, R = 3.25 km h⁻¹. Fires were burning under similar weather conditions, with a Grassland Fire Danger Index of: top = 23; middle = 26; bottom = 26.

The 10-m open wind speed, standing fuel moisture content, GFDI and rate of fire spread were found to be normally distributed according to the Shapiro–Wilk test of normality (P > 0.05). Paired t-tests for differences between control–treatment pairs found significant differences in rate of fire spread, with the Unharvested crop condition fires spreading on average 2.1 km h⁻¹ faster than the fires in the Harvested condition (Table 4). Similar results were obtained for the comparison between the Baled and Harvested condition fires, with the fires spreading on average 2.1 km h⁻¹ faster in the taller fuel condition. Nonetheless, this last comparison is not strictly within the Fr <0.5 flow regime of MSM18 and can be discounted from the analysis.

Table 4.  Differences in environmental and fire behaviour variables for paired burns (treatment – control; t-test) in Triticum spp. in Cruz et al. (2020)
Differences are given in the units of the variable. GFDI, Grassland Fire Danger Index

The results of our analyses of experimental fire spread data in field conditions are clear and unequivocal. For the range of fuel bed heights typical of temperate grasslands (e.g. Sneeuwjagt and Frandsen 1977; Anderson 1982; Andrews et al. 2006) and Fr constraints used by MSM18 (Fr <0.5), a reduction in grass height, with or without removal of the cut fuel (i.e. with or without change in bulk density), led to a corresponding reduction in average rate of fire spread. This is the opposite of the results presented by MSM18, in which a strong and inverse effect was determined. Cohen et al. (2006) reasoned that increases in fuel depth allow for the extension of the flow length of potential flame contact with unburned fuel, with fuel bed depth being directly related to potential convective heat transfer for flame spread. A reduction of fuel bed height would reduce the flame potential flow length, and consequently reduce the efficiency of convective heat transfer to fuels in the path of the advancing fire.

We disagree with the statement by MSM18 that the dependence of rate of fire spread on grass height is a matter of debate. The results obtained by Cheney et al. (1993) on a broader analysis than the one presented here showed a direct and significant correlation between fuel bed height in grasslands and rate of fire spread (r = 0.23, P < 0.01). Cruz et al. (2016, 2018) failed to find such a relationship because the data in these studies originated from a variety of structurally different grasslands and autocorrelations between fuel height and fuel load or fuel height and fuel particle surface area to volume ratio masked the effect of fuel height. For example, when considering unmodified grasses, finer grasses with higher surface area to volume ratios tend to be shorter. In contrast, the high fuel load grasses in the Cruz et al. (2018) study were taller and coarser. In the Cruz et al. (2018) dataset, fuel load showed a stronger relationship with rate of fire spread (r = −0.50; P < 0.001) than fuel height and showed a reduction in the rate of fire spread for values above ~0.6 kg m⁻². Given these autocorrelation issues, one should not attempt to derive a statistical secondary effect of fuel height on rate of fire spread from the datasets of Cruz et al. (2016, 2018). With the data used in the present comment, we removed any confounding effects of fuel coarseness from the analysis and thus can directly replicate the simulation assumptions in MSM18.

The effect of grass height observed in the Cheney et al. (1993) and Cruz et al. (2020) subsets of data are distinct in magnitude. This is not surprising because the observed reduction in rate of fire spread depends on other structural characteristics of the fuel bed, including the absolute reduction in fuel height and associated burning conditions, namely the wind speeds, under which the fires were conducted.

A reduction in fuel bed height while maintaining bulk density, as simulated by MSM18, results in a corresponding decrease in the fuel available for combustion. It is not easy to disentangle the effect of these two fuel characteristics on the rate of fire spread, but experimental results in grasslands (Cheney et al. 1993), forest fuels (McCaw et al. 2012) and shrublands (Anderson et al. 2015) suggest that fuel bed height has a stronger effect than fuel load. These findings are applicable when fuel loads are not limiting in regard to the formation or oxidation of gas-phase volatiles that we see as flame.

Without a detailed forensic investigation of the physics-based model’s formulations, assumptions and settings, it is impossible to isolate any one reason as to why MSM18 obtained the results they did. However, several candidate factors may be identified from the detail provided in their paper. Foremost among these could be the representation of the fuel as a homogeneous layer (i.e. boundary fuel) and the inherent assumptions regarding the effect on drag in air flow over the fuel bed. These assumptions include that heat is released from above the fuel bed rather than within the fuel bed, and that grass blades were modelled solid cylinders (with commensurate drag coefficient) that are not malleable to bend with the wind, thus affecting roughness length (Penman and Long 1960), which might perhaps mean the drag force imposed by the fuel on the air flow over it is overstated. The result of this could be an underprediction of the dynamic forces and an overprediction of the buoyant forces at the flame front.

We commend the efforts of MSM18 to attempt to derive an understanding of the effect of fuel characteristics on fire behaviour through numerical simulations using a physics-based model. Such approaches are necessary to improve model behaviour and ultimately an understanding of fire dynamics. Nonetheless, awareness of existing empirical evidence, namely from well controlled field experiments, is key to evaluating the realism of numerical experiments and is ultimately necessary if one aims to use physics-based models to investigate still unanswered questions on the dynamic nature of wildfire propagation and particularly the response of fire to changes in fuel structure. There are several published results and datasets available from studies on grass fire propagation that were not mentioned by MSM18, in particular those from Cheney et al. (1993), a large-scale experimental study aimed at understanding the effect of grass fuel load and height on fire behaviour. There are also other relevant published data available, such as in Sneeuwjagt and Frandsen (1977) and Clark (1983), where one can investigate the effect of grass fuel bed height on rate of fire spread. Awareness of these datasets and the empirical trends would have allowed MSM18 to understand the incongruencies in their simulation results, the likely causes of it, and allow for improvements in the model understanding and future behaviour.

Often modellers aim to publish model evaluation studies where it is shown how well a model fits the real-world data. This is not difficult to achieve with fire behaviour models. Given the degrees of freedom in a physics-based model of fire behaviour, one can always fit a model to an experiment. Given the unknowns in the physical processes driving fire propagation, the uncertainty inherent in any fire quantity measurement, either indoors or outdoors, and the coarseness of existent numerical models, we believe that modellers should not attempt to evaluate or ‘validate’ a model by showing how well the model fits reality, but strive to show how models fail to replicate observed behaviour (Watts 1987). Failure drives improvement. Investigating a model’s poor or erroneous results will contribute to its improvement to a level where it may perform adequately over a broad range of conditions and, ultimately, be able to be used to contribute to successful fire management.

The authors declare that they have no conflicts of interest.