Incorporating burn heterogeneity with fuel load estimates may improve fire behaviour predictions in south-east Australian eucalypt forest

Study area

The study was undertaken in eucalypt forests within the state of New South Wales (NSW), in south-eastern Australia (Fig. 1). The climate is temperate, with warm summers and cold winters, with no dry season (Bureau of Meteorology 2022b). Mean annual rainfall ranges from 600 to 1500 mm, increasing towards the coast, and declining inland of the Great Dividing Range (a series of mountain ranges running approximately parallel with the coast; Bureau of Meteorology 2022b). Average minimum winter temperatures range from 0 to 9°C and average maximum summer temperatures range from 21 to 30°C (Bureau of Meteorology 2022b). Temperatures are coolest in the high elevation areas along the Great Dividing Range, and inland of the Range. Eucalypt forests dominate the native vegetation of the study area, and range from wet to dry sclerophyll forest types (Keith 2004). Ecologically sustainable fire intervals for these forests are estimated at 5–50 years for the dry sclerophyll forests and 25–60 years for the wet sclerophyll forests, with crown fires to be avoided at the lower range of the interval (Kenny et al. 2004).

Fig. 1.

(a) Location of study plots for fuel load observations and the case study 695 wildfire. (b) Fire severity of the 695 wildfire and mapped perimeter of the preceding hazard reduction burn. (c) Fire severity of the hazard reduction burn. Fire severity mapped following Gibson et al. (2020). Refer Fig. 2 to clearly see the intersection of the 695 wildfire and prior hazard reduction burn.

Changes in fuel loads with increasing fire severity

Post-fire fuel loads (t ha⁻¹) were measured approximately 1 year following the 2019–20 Black Summer wildfires. Sampling at 1 year post-fire allowed for the influx of scorched foliage and branches from the canopy to the surface fuel layer, which can occur over the weeks and months following fire (Price et al. 2022). We measured loadings within four forest types: (1) western slopes dry sclerophyll; (2) Sydney hinterland dry sclerophyll; (3) south-east dry sclerophyll; and (4) Southern Tableland wet sclerophyll forests. Forest type classifications were based on vegetation classes described in Keith (2004). These forest types were selected because they span a relatively broad climatic gradient (670–1088 mm mean annual rainfall; Bureau of Meteorology 2022a), exhibit varying surface fine fuel loads (Watson 2012), and they were subject to a range of fire severities during the 2019–20 fires, but some unburnt forest remained in close proximity to burnt forest. The study design is described in greater detail in Nolan et al. (2022), which presents observations of post-fire tree mortality and above-ground carbon stocks

We measured 15 plots in each forest type spanning the range of fire severity classes. Fire severity was classified in the field into one of five categories following the severity class definitions of Gibson et al. (2020), and based on the proportion of leaf scorch and consumption in understorey and overstorey strata. Fire severity was assessed from satellite images, and confirmed in the field. Severity classes were: unburnt, low (<10% canopy scorch), moderate (20–90% canopy scorch), high (>90% canopy scorch but <50% canopy consumption) and extreme (>50% canopy consumption). We measured three unburnt plots for each forest type. We similarly aimed to measure three plots in each of the other fire severity classes, but this was not possible due to the nature of the wildfires where few locations burnt at both low and extreme severity. Due to the low replication in each fire severity category for some forest types (n = 2), we aggregated our plots into the following two fire severity categories: (1) ‘low-moderate’ fire severity, which represents predominantly understorey fire; (2) and ‘high-extreme’ fire severity, which represents homogenous canopy fire, n = 4–8 for these aggregated categories within each forest type.

Burn heterogeneity effects on subsequent fire spread

Case study fire

Our case-study wildfire was the ‘695’ fire, which burnt 19,090 ha in February 2018 (Fig. 1b). The 695 fire progressed over multiple days. We selected this fire as our case-study because a prior hazard reduction burn (conducted in June 2016; Figs 1c, 2), was expected to slow or limit the spread of the 695 wildfire, but this did not occur. While the area of the mapped hazard reduction burn was 10,172 ha, our analysis of fire severity found over 80% of the mapped burn area remained unburnt (see below for details on fire severity mapping). Where it did burn, the hazard reduction burn was primarily an understorey fire of low to moderate severity (Fig. 1c).

Fig. 2.

Fire history within the study area. (a) Fire history using current burned area mapping, and (b) fire history modified by fire severity mapping to account for unburnt patches within fire perimeters.

The 695 fire burnt predominately within Wollemi National Park, which is a 501,000 ha park located 130 km north-west of Sydney. Mean annual rainfall is estimated at 812 mm, and mean monthly minimum temperatures range from 3 to 14°C and mean monthly maximum temperatures range from 13 to 28°C. Climate data extracted from the SILO Australian Climate Dataset (http://www.longpaddock.qld.gov.au/silo; Jeffrey et al. 2001). Vegetation reflects that found in the wider study area, i.e. includes wet and dry sclerophyll forest types.

Fire history mapping was obtained from the New South Wales Rural Fire Service. Fire severity of the wildfire and hazard reduction fire, as well as all observed fire dating to the 1989–1990 fire year, was mapped with satellite imagery following Gibson et al. (2020), which uses an extensively trained and tested Random Forest supervised classification algorithm applied to satellite imagery from before and after the fire (Gibson et al. 2020; Department of Planning Industry and Environment 2021). The fire severity mapping approach in Gibson et al. (2020) classifies unburnt and extreme fire severity with very high accuracy (mean balanced accuracy statistic >0.95), low and high fire severity with high accuracy (mean >0.80) and moderate fire severity with relatively lower accuracy (mean >0.70).

The hazard reduction fire was mapped using Landsat satellite imagery (30 m resolution), whereas the wildfire was mapped using the higher resolution sensor Sentinel 2 (10 m) which was launched between the two fire events. The combined use of Landsat and Sentinel 2 sensors in remote sensing applications for fire extent and severity mapping has demonstrated substantial agreement between the outputs from the different sensors (White and Gibson 2022).

Changes in fuel loads with fire severity

Fire reduced fuel loads across all strata examined (Fig. 3); however, there was generally no change in fuel loads with increasing fire severity, with the exception of bark fuel loads (Fig. 3, Table S1). At 1 year post-fire, surface fuel loads ranged from 12.6 to 25.2 t ha⁻¹ in unburnt forest, with fire reducing these fuel loads to 2.6–5.6 t ha⁻¹. The surface fuels were primarily comprised of litter fuels, which ranged from 10.5 to 25.1 t ha⁻¹ in unburnt forest, with near surface fuels only ranging from 0.1 to 2.4 t ha⁻¹ (Fig. S1). The wet sclerophyll forest was the only forest type where higher fire severity was associated with significantly lower fuel loads at 1 year post-fire. For this forest type surface fuel loads following low-moderate severity fire were 5.3 t ha⁻¹ and following high-extreme severity fire was 2.6 t ha⁻¹.

Fig. 3.

Variation in fine fuel loads as a function of fire severity. Fuel strata represented are (a–d) surface fuels (i.e. litter and near surface fuel strata combined), (e–h) elevated fuels and (i–l) bark fuels. Data illustrated are averages ±1 s.e. Differing letters above bars indicate significant differences among fire severity classes (P < 0.05). For elevated fuels, lower case letters represent differences among fire severity categories at 1 year post-fire, while upper cases letters represent differences among fire severity classes excluding post-fire regrowth (i.e. the approximate fuel load immediately post-fire).

Elevated fine fuel loads ranged from 0.1 to 1.5 t ha⁻¹ in unburnt forest (Fig. 3). Notably, there was large variation in elevated fuel loads within unburnt forest, as evidenced by high standard errors, which ranged from 0.1 to 1.1 t ha⁻¹. At 1 year post-fire elevated fuels did not differ significantly between burnt and unburnt forest for any of the forest types. However, when the post-fire re-growth was excluded, which provides an approximation of immediate post-fire fuel loads, we did observe significantly lower fuel loads in burnt forest for two of the four forest types. Again, there was no change with fire severity.

Bark fine fuels were estimated to range widely (2–4 t ha⁻¹) in unburnt forest (Fig. 3). Forest burnt at high-extreme fire severity had significantly lower bark fuel loads compared to unburnt forest, estimated at 1 t ha⁻¹ for all four forest types. Forest burnt at low-moderate fire severity had intermediate fuel loads (1.8–3 t ha⁻¹), which did not differ significantly from unburnt forest.

Fuel hazard scores showed similar trends to the fuel load data (Fig. S2, Table S2). That is, litter hazard scores were significantly lower in burnt compared to unburnt forest, but there was no effect of the severity of the fire. There was no significant effect of fire for near surface hazard scores for any of the forest types, with the exception of the Southern tableland wet sclerophyll forest, where the near surface hazard scores were significantly lower in forest burnt at high-extreme fire severity compared to unburnt forest. For elevated hazard scores, despite a trend of lower scores for burnt compared to unburnt forest, there was no significant differences observed, except for the Western slopes dry sclerophyll forest. For this forest type, forest burnt at low-moderate fire severity had significantly lower hazard scores compared to unburnt forest. Finally, for bark hazard scores, there were significantly lower values for forest burnt at high-extreme fire severity, but no difference between unburnt forest, and forest burnt at low-moderate severity.

Adjusted fuel accumulation curves

Modified fuel accumulation curves for surface fuels mostly aligned with existing curves (Fig. 4a–d), indicating that the steady-state and initial fuel load parameters in the current curves were similar to our field-based observations. The exception was the South-east dry sclerophyll forest type, where there were large differences between the current and modified curves, due to differences in the estimates of steady-state fuels. The current curves assume steady-state surface fuels are 12.0 t ha⁻¹, while we observed fuel loads of 25.2 t ha⁻¹ (Table S3).

Fig. 4.

Fuel accumulation curves for surface (a–d), elevated (e–h) and bark (i–l) fine fuels. Represented are the current fuel accumulation curves used operationally (‘Original model’), in addition to modified curves based on our observations of fuel loads in long unburnt forest and at 1 year post-fire. Note, where we did not observe a significant effect of fire severity on post-fire fuel loads, we pooled observations across fire severity classes to estimate the initial fuel load. Where we observed no significant effect of fire on fuel loads, we did not model fuel accumulation, instead assuming fuels remained at steady-state conditions.

In contrast to surface fuels, there were considerable variations between current and modified curves for elevated fuels (Fig. 4e–h). This was primarily due to our estimates of steady-state fuels (0.1–1.5 t ha⁻¹) being much lower than those in current curves (2.0–5.0 t ha⁻¹). Our estimates of elevated fuel loads over time-since-fire also differed for two of the four forest types, where we did not observe significant differences in initial post-fire fuel loads in elevated fuels, and thus did not model fuel accumulation over time since fire.

For bark fuels, our modified fuel accumulation curves for high-extreme fire severity were either lower or higher than current operational curves (Fig. 4e–f). The differences between our curves and the operational curves reflect our differing observations of steady-state fuel loads. For example, steady-state bark fuel load estimates were considerably higher for the Western slopes dry sclerophyll forest in our modified curves (3.3 t ha⁻¹), compared to current curves (2 t ha⁻¹), resulting in much higher estimated fuel loads over time since fire. For forests burnt at low-moderate fire severity, where we did not observe a significant difference between burnt and unburnt forest, we found no change in fuel loads over time since fire, and thus our estimates of bark fuel loads were much higher in the early post-fire years compared with the current curves.

Burn heterogeneity effects on subsequent fire spread

The first PHOENIX simulation (‘current operational’ simulation), predicted that the spread of the case study 695 wildfire would be limited by the hazard reduction burn conducted in the preceding 2 years (Fig. 5a). In contrast, the PHOENIX simulation that accounted for the unburnt patches within historical fire perimeters (the ‘burn heterogeneity’ simulation, Fig. 5b) resulted in a much larger predicted burned area. For this simulation, the prior hazard reduction burn was predicted to have no effect on the spread of the wildfire (Fig. 5b). For the third simulation (‘Burn heterogeneity and bark’ simulation, Fig. 5c), we additionally modified bark fuel accumulation curves such that bark fine fuels were not modelled to decline following understorey fire, based on our results (Fig. 4). We additionally modified bark fuel accumulation curves based on our observations of steady state fuel loads. This third simulation predicted a burned area that was most similar to the observed burned area of the 695 wildfire (Fig. 5d).

Fig. 5.

PHOENIX simulations: (a) current operational; (b) ‘burn heterogeneity’ – fire history modified by fire severity mapping to account for unburnt patches within fire perimeters; and (c) ‘Burn heterogeneity and bark’, same as (b), in addition to bark fuel accumulation curves being modified, based on field survey data.

For all three PHOENIX simulations, predicted flame height, fireline intensity and rate of spread were all similar (Fig. 6). However, the predicted spotting distance for the third simulation (median value of 469 m) was lower compared to the second simulation (median value of 674 m), reflecting the lower predicted bark fuel loads used as inputs in this simulation. This lower spotting distance resulted in a smaller burned area compared to the second simulation (15,348 ha compared to 32,441 ha). The simulation that was most similar to the actual burn area was the third simulation (Jaccard index of 0.51, compared to 0.30–0.38; Table 1).

Fig. 6.

Boxplot summaries of fire behaviour from PHOENIX simulations, including (a) area burnt, (b) flame height, (c) fireline intensity, (d) rate of spread, and (e) spotting distance. The simulations are: (1) current operational; (2) ‘burn heterogeneity’ – fire history modified by fire severity mapping to account for unburnt patches within fire perimeters; and (3) ‘Burn heterogeneity and bark’, same as (b), in addition to bark fuel accumulation curves being modified, based on field survey data.

Changes in fuel loads with fire severity

Similar to this study, Price et al. (2022) found post-fire surface fuel loads did not differ due to fire severity. Price et al. (2022) attributed this to post-fire litter deposits increasing with fire severity, offsetting the increased consumption of surface fuels. Surprisingly, we also found no significant effect of differing fire severity on elevated fuels at 1 year post-fire, as both understorey and canopy fire similarly reduce elevated fuel loads. This similarity in elevated fuel loads may reflect the smaller impacts of fire on the understorey in forest burnt at low to moderate fire severity, and rapid recovery of understorey vegetation in forest burnt at high to extreme severity in the year after the fire. In forest burnt at high to extreme severity, the high-light environment facilitates prolific post-fire recruitment in eucalypt forests (Bennett et al. 2016). We also found no effect of fire severity on the shrub density (Table S5). Other studies found elevated fuel loads, or shrub density and cover, are affected by fire severity (e.g. Landesmann et al. 2021; Barker et al. 2022). These studies generally examine elevated fuel trajectories over time or at a later time period than our study. These studies argue that increasing fire severity creates an environment with high light and reduced competition in the understorey while the canopy recovers (Bennett et al. 2016; Gordon et al. 2017; Etchells et al. 2020) and in turn this promotes increased shrub growth, and hence increased elevated fuel loads.

We observed spatial variation in fuel loads, particularly for elevated fuels. For example, in south-east dry sclerophyll forest, elevated fine fuel loads in unburnt forest ranged from 0.1 to 3.8 t ha⁻¹ (Fig. 3f). Large spatial variability in elevated fuels has similarly been observed in south-east Australian eucalypt forests (McColl-Gausden et al. 2020; Nolan et al. 2022). There was less spatial variation in surface fine fuel loads when compared to elevated fuels (Fig. 3). The exception was the south-east dry sclerophyll forest type where our surface fuel load observations in unburnt forest were much higher than those modelled in current operational fuel curves (Fig. 4b). This may have been because the study plots were transitional between the dry sclerophyll forest and the nearby wet sclerophyll forest, which has much higher fuel loads. Spatial variation in fuels within vegetation communities is partially due to historical variation in fire frequency, severity and seasonality, which all affect plant recruitment (Enright et al. 2015; Miller et al. 2019). Additionally, factors that affect species distributions, such as fine-scale variation in soil attributes and topography, likely affect fuel loads. Thus, fuel load models may be improved by modelling fuels as a function of biophysical parameters that vary at a finer spatial resolution than vegetation communities. Indeed, recent biophysical models of fuel hazard scores have found that climatic and edaphic factors are good predictors of fuel hazard scores (Jenkins et al. 2020; McColl-Gausden et al. 2020).

Adjusted fuel accumulation curves

Our modified bark fuel accumulation curves following canopy fire were either lower or higher than current operational curves, reflecting differences in steady-state fuel loads. For the western slopes dry sclerophyll forest, we observed higher bark fine fuel loads in unburnt forest than predicted by current fuel curves. This variation from the fuel curve was due to the presence of species which produce higher bark fuels than the ironbark species which typically characterise this vegetation type (Hines et al. 2010; Horsey and Watson 2012). For the other three forest types, we observed lower bark fine fuel loads in both unburnt and burnt forest than predicted by current fuel curves. These results also reflect differences in the presence of species that produce high bark fuel loads, e.g. stringybark species, which produce the highest fine fuel loads (Hines et al. 2010).

Burn heterogeneity effects on subsequent fire spread

Accounting for burn heterogeneity in fuel load predictions was found to improve fire spread simulations from PHOENIX for our case study 695 wildfire (Fig. 5, Table 1). A previous hazard reduction burn had been predicted to limit the spread of the 695 wildfire. However, when the large area of unburnt patches within prior fire perimeters was incorporated into fire history mapping, the previous hazard reduction burn was no longer predicted to limit the spread of the wildfire. Thus, fine-scale resolution burned area mapping has the potential to improve fire behaviour simulations. As demonstrated here, utilising fire severity maps to improve fire history mapping substantially improves fuel load estimates, and thus predicted fire behaviour. Further improvements in predicted fire spread were gained by modifying bark fuel accumulation curves based on field observations (Fig. 5c). Eucalypt bark produces firebrands that can travel large distances, igniting new fires (McArthur 1967; Penman et al. 2017). Thus, improved models of fine fuel loads, particularly bark fuels that affect spotting distance, can further improve predictions of fire spread. However, we note that improved fuel load mapping may not necessarily lead to improved fire behaviour simulations under all weather conditions, with fuel exerting less influence on fire behaviour under more extreme fire weather conditions (Clarke et al. 2022).

While our results suggest that patchy hazard reduction burns may not be as effective as those with a homogenous burned area in reducing fire spread under wildfire conditions, it is important to note that fire refugia are important for supporting biodiversity and maintaining ecosystem processes (Cawson et al. 2013; Krawchuk et al. 2020). Indeed, many planned burns, including hazard reduction burns, have the aim of creating a mosaic of burnt and unburnt areas (Duncan et al. 2015). Our results also suggest that high severity fire is most effective at reducing bark fuel loads; however, it is important to note that our study does not consider longer-term vegetation dynamics that may affect fuel load and structure, and subsequent fire behaviour, that may operationally influence the ability to contain a fire or result in fires escaping. For example, a study by Barker et al. (2022) in the same region found that at 5 years post-fire, areas burnt at high severity had greater elevated fuel loads and fuel connectivity compared to areas burnt at low severity. In a related study, Barker and Price (2018) found that crown fire was more than twice as likely following a previous crown fire, compared to a previous understory fire. Similarly, in a study on mixed conifer forest in the USA, Coppoletta et al. (2016) found that the effects of high severity fire on standing deadwood and shrubs promoted high severity fire in subsequent fires. We also note that Barker and Price (2018) found that the lowest likelihood of high severity fire occurred in the longest unburnt forest, which indicates that the shape of fuel accumulation curves may need to be modified, particularly for elevated fuels (Zazali et al. 2021).