Comment on ‘Self-thinning forest understoreys reduce wildfire risk, even in a warming climate’

In this comment we examine a recent study published in Environmental Research Letters that analysed fire history data from forests in Western Australia to suggest that changes in forest structure result in a long-term reduction of fire risk after 56 years since last fire. We examine the data underpinning this study and find that its strongly skewed sample size distribution creates a bias to the extent that the analytical approach would find a pattern of declining fire risk even when there was no decline. Moreover, the very small sample sizes of the longest unburned forests mean that fire mapping errors as small as 1–2 ha can reverse key findings. With documented mapping errors orders of magnitude larger, the dataset is not robust to analysis at this level of precision. An appropriate conclusion, taking into account these detection and sensitivity issues, would be that likelihood of subsequent wildfire is reduced in the first ∼6 years following fire, and remains fairly consistent at a higher level for at least the next 3 decades, with no evidence for a long-term reduction of fire risk. This is relevant given that many fire and forest management decisions are made based on scientific literature. Rather than wildfire risk reducing with increasing time since fire, our projections indicate that ceasing active fire management in the sampled forests could result in landscape wildfire extent 25%–65% above current levels. We recommend further steps that would help provide sound, evidence-based knowledge to inform science, management, and policy.


Introduction
Recently, Zylstra et al (2022a) analysed 55 years of Western Australian fire history data to assess the likelihood of forest wildfire with increasing time since fire (TSF), in a paper titled 'Self-thinning forest understoreys reduce wildfire risk, even in a warming climate.' This study reported that the likelihood of forests experiencing wildfire was relatively low for the first 5-7 years after fire, higher between 8 and 56 years, and reduced again between 57 and 64 years after fire.From these results, the authors argue that the absence of fire leads to a 'long term reduction of fire risk' and suggest that active fuel management in long unburned forests is counter-productive and should be avoided.Such a proposal to limit fuel management in southwestern Australia's strongly seasonal, fire-prone landscape has important implications for a long-established policy framework, public safety, and the ecological paradigm informing land and fire management.
We consider that (2022) contains substantial and important problems in their analysis, as well as serious oversights in study design and interpretation of existing knowledge around fire hazard and fuel dynamics in forests of southwestern Australia.To support these statements, we present a detailed assessment of the data, analysis and inferential limitations, and correct several misconceptions about the effect of TSF on fuel structure and fire behaviour.We caution against using the results presented by Zylstra et al (2022a), considering them broadly unreliable and their conclusions unsupported.We conclude with a series of research questions and approaches that would help further develop sound, evidence-based knowledge to inform fire-related science, management, and policy.

Data structure and inference
Zylstra et al (2022a) do not use direct observation of forest and fuel structure dynamics, or fire behaviour to assess self-thinning of forest understoreys, or wildfire risk, or their relationship.Instead, they infer this relationship via analysis of fire frequency patterns in a publicly available spatial database of fire history for Western Australia's public lands (DBCA Fire History 'DBCA-060'; DBCA 2018).The authors restricted these data to mapped forest vegetation in National Parks, resulting in a widely dispersed but geographically unbalanced sample of 528 343 ha, comprising <20% of forests in southwestern Australia.
Key to the data structure are: 'sample years' and TSF.TSF is, for every year observed between 1964 and 2018, the number of years that have elapsed in an area of forest since it was last recorded as experiencing prescribed burning or wildfire:(2022) called this 'forest age'9 .'Sample years' are years between 1964 and 2018 in which different samples of forest with a given TSF were observed.For instance, an area of forest that burned in both 2000 and 2010, had a TSF of 1 in both 2001and 2011and a TSF of 2 in 2002and 2012, etc, so between 2001and 2018, TSF 1-8 have 2 sample years and TSF 9 and 10 have one sample year, with fire occurring only in one sample year and TSF (TSF10 in 2010).
For each year between 1964 and 2018, the authors extracted the total area of forest in each TSF (commencing from 1955), and the total area of forest in each TSF that experienced wildfire in a 1 -ha spatial grid.The authors define 'flammability' (notated 'f ' here) of each TSF as the mean proportion of forest area within each TSF that experienced wildfire, averaged across sampled years i.e.
where n is the number of sample years observed for the TSF, a and b are the area of forest, and the area of forest that burned in wildfire, respectively, in the TSF in sample year i, (see supplementary information for further details), with TSF ranging from 1 to 64 years.The premise of this approach is that the average annual proportion of forest area in each TSF that experienced wildfire (i.e.f ) represents the annual likelihood that forests of that TSF will burn if exposed to ignition or adjacent fire, so that change in f with TSF is change in wildfire likelihood with TSF.This approach does not consider patterns of when or where ignitions or adjacent fire occurred, or the spatial arrangement of forests or of areas with different TSF in the landscape-when in fact the likelihood that an area will experience wildfire is closely related to patterns of ignition sources and density, and effectiveness of suppression response (Cary et al 2009, Plucinski et al 2014, Collins Kathryn et al 2015).Rather, the authors' premise assumes that patterns of fire likelihood with TSF can be detected given a very large sample size of forest area, number of fire ignitions, and opportunities for spread into areas of different TSF.While the dataset is large and potentially robust in some respects (more so among the shortest TSF), it is extremely skewed such that the areas of long unburned forest key to testing hypotheses have a critically small sample size.

Low sample sizes lack power to support wildfire risk conclusions
The conclusions reached by Zylstra et al (2022a) that fire likelihoods decline as forests age hinge upon samples with the longest TSF (i.e.56-64 years).However, this portion of the dataset is very poorly sampled in terms of both number of sample years and forest area.This is an emergent property of the dataset.For instance, TSF 64 has just one sample year because only one 64 year period fits in the 1954-2018 record period, these represent forest area mapped as burning in 1954 (the first year of the record) that remained unburned up to 2018, the final year of the record when it was observed as TSF 64.Two sample years exist for TSF 63 as two 63 year periods can be sampled with forests last burned in 1955 and 1956 sampled in 2017 and 2018, and TSF 62 has 3 sample years, etc.With 3, 2 and 1 sample years respectively, there are only six opportunities to assess the proportion of wildfire occurring in the three TSF 62-64.A patch of forest burned in the 1950s would need to remain unburnt during the entire sampling period to be able to be sampled over 60 years later, and this is rare given the highly fire-prone nature of the region and extent of planned fire-the forest area sampled in the three TSF 62-64 range from 8 to 27 hectares.In contrast, TSF 1 has 55 sample years, as every year in the record included opportunity to sample forests one year after fire, yielding 3.1 million ha of sampled forest.The total area of forest sampled declines over five orders of magnitude (figure 1(A)).
Published DBCA-060 metadata (DBCA 2021) state that fire polygon boundary accuracy varies, with many classes of 'capture method' listed in the dataset.In the study period and area relevant to Zylstra et al (2022a), >95% of prescribed burns, and 80% of wildfires were mapped through methods that have stated accuracy of ±500 m or ±1 km (most being 'Historic maps digitized (±500 m)').Just 8% of wildfire area was mapped using methods (mostly 'Remote Sensing, spatial resolution 25 meters (Landsat)') with accuracy stated as <±100 m.Zylstra et al (2022a) extract data in a 1 ha spatial grid (these methods are not detailed, but forest and fire area data are recorded at integer hectare precision), while the most common mapping accuracy of the dataset would imply ±50 ha for every 1 km of fire boundary.
Small sample size interacts with the low overall frequency of wildfire in this dataset to overwhelmingly bias estimates of 'flammability' (f in figure 1).Averaging the f scores across the 64 different TSF (f mean (all) = 0.0205; figure 1(B)) suggests a mean wildfire likelihood across TSF of ∼2% per year (see supplementary information for further details).As forest and fire area is recorded at 1 ha precision, the smallest sample area where the average amount of fire (f mean (all) ; ∼2%) can be detected is 50 ha-because 2% of 50 ha is 1 ha, and areas of fire <1 ha are not recorded.If fire impacted 2% of a 25 ha forest sample, the burned area would be 0.5 ha, and, as this is smaller than 1 ha, no fire would be recorded: the 25 ha sample would report a value of 0% burned even though the amount was exactly the same as the overall forest average.In fact, samples <50 ha will always report 0% fire unless fire extent exceeds the forest average, and fire extent must at least double this (i.e.>4%) if samples <25 ha are to report any fire at all.
Zylstra et al (2022a) derive their findings from ttests that compare the average f for all TSF longer than each tested TSF to the average f of all shorter TSF.If many samples were smaller than the 50 ha detection limit for average f, and these were concentrated at one extreme of TSF, then this underreporting of fire would bias results.
In the analysed data, 89 samples are <50 ha (figure 1(C)), these occur in all 27 of the TSF >37 years (figure 1(C)), 33 of these 89 samples are <25 ha.The presence of these spatially small samples means that all of the longest TSF under-detect wildfire.Crucially, the three longest TSF have no samples >50 ha (max = 27 ha).Consequently, those TSF would detect no fire, and report a flammability of 0%, even if they did experience the average amount of fire; figure 1(C)).Due to this bias, average fire extent across all eight mature TSF must be nearly double f mean (all) before calculated f mean (mature) even reaches f mean (all) .This problem invalidates the main finding and conclusion of Zylstra et al (2022a): the sampled data array finds declining wildfire risk with TSF, but would do so even if wildfire likelihood actually increased.It literally cannot determine whether flammability declines or not.Because the data will always report declining wildfire risk in the longest TSF whether risk declined or not, the analysis of Zylstra et al (2022a) does not provide evidence that wildfire risk declines with age.The finding showing that it does is an artefact of spatial size of samples used.
As sample sizes fall below detection limits, small changes in measured fire area have increasing influence on results.As such, the assessed wildfire risk patterns are supremely sensitive to error in fire mapping in longer TSF.In the longest TSF assessed, the single year sampled had a forest area of 27 ha, with no fire detected, giving a f 64 of 0. If one ha of fire had occurred, f 64 would be 3.7% (well over the mean f, i.e. not a decline), this represents both the minimum f detectable in this TSF, and the sensitivity of the TSF to 1 ha error.Across the 8 longest ('mature') TSF, the average wildfire detectability threshold (minimum non-zero f detectable with 1 ha resolution), and sensitivity to a 1 ha mapping error (change in f resulting from 1 ha error) range from f = 0.5% to 3.7% (figure 1(B); average = 1.6%).These are calculated by adding 1 ha to each sample in each TSF in turn.This level of sensitivity is comparable to the difference between f mean (all) and f mean (mature) (1.7%).One or two hectare-scale mapping errors can cause critical (t-test) results to flip between significant and non-significant.Not taking this sensitivity into account would lead to invalid conclusions.Errors of this scale are inherent in the type of data analysed, and well below the stated level of reliability for the dataset.
As so many critical sample sizes are small, and wildfire is relatively infrequent, 1 ha and smaller scale variation in fire extent has consequences for detection limits and sensitivity in the Zylstra et al (2022a) analysis.The 1955-2018 fire history dataset analysed has substantial limitations at scales much coarser than 1 ha.Hamilton et al (2009) reviewed spatial accuracy of a large part of this fire history dataset, finding that, prior to 1972, no areas of mapping had high confidence and 70%-80% had low or very low confidence.Only 20% of area mapped after 1972 had high confidence.More recently, Dixon et al (2022) report that 22% of area mapped as experiencing fire between 2005 and 2020 in the same spatial dataset, was in fact not burned.Curation and handling errors can also be significant.A single mapping error associated with the largest wildfire in the dataset accounts for 28 192 ha of fire, 6.1% of all wildfire sampled.This error arose when the fire was mapped using two different techniques, one correctly attributing it as occurring in March 2003, and the second incorrectly attributing the fire to the subsequent year.This duplication implied that a vast area of fire in burned through forests one year after it previously burned (i.e.TSF = 1).Corrected in DBCA 060 after its discovery in 2020, this error remains in the Zylstra et al (2022a) analysis of 2018 data.It is apparent where they report wildfire risk of TSF = 1 (f 1 ) to be 0.63%.Its correction results in an f 1 of 0.10% and changes the mean wildfire risk across all TSF (f mean (all) ) to 2.04% from 2.05% (figure 1(B)).Emphasising the skew in sample area, 28 192 ha is larger than the total area of fire reported in the 42 TSF classes greater than TSF = 22.

Exploring management consequences of declining wildfire risk in longer TSF
The paper's conclusions on forest fire risk have been used by the authors and others to argue for changes in fuel management policy.Zylstra et al (2022b; The Conversation) argue that fuel reduction burning may be counter-productive because it maintains forests in a state of high wildfire risk, whereas preventing fire and ceasing prescribed burning will allow forest understoreys to mature to a condition where they become less likely to burn.It is possible to assess this policy of fire exclusion, using the same data, by projecting the reported average proportion of forest area of each TSF that experienced wildfire (i.e. the f for each TSF) on to the current distribution of forest area in each TSF to calculate the total area of wildfire, and the resulting distribution of forest area in each TSF in the next, and (by iteration) subsequent year(s).This process projects changes in the total area of wildfire into the future (figure 1(D)).Because prescribed burning is prevalent, and effective for up to seven years in this dataset, ceasing its application across all forest ages results in a rapid increase in projected area of wildfire over eight years (figure 1(D)).Average wildfire extent would then remain elevated in a band between 25 and 65% above current levels for a further 55 years.Projected wildfire extent drops thereafter in this model but not to below 2018 wildfire levels until 75 years.This model input includes the detectability and sensitivity problems outlined above.Assessing this, projections applying the average likelihood of fire across all forest ages (f = 2.04) instead of the reported average of 'mature' forests (f = 0.31) for fire likelihood in forests with TSF > 64 years, show that average annual wildfire extent is elevated for six decades, stabilising at 36% above 2018 levels (figure 1(D)).

Evidence on fuel dynamics and self-thinning with TSF: fire behaviour consequences
Zylstra et al (2022a) postulate the likelihood a wildfire will burn an area is largely a function of the understorey vegetation.Presenting no new evidence on self-thinning, they claim that declining wildfire risk occurs as the understorey layer self-thins with TSF.Although some aspects of fire behaviour are influenced by the structure and arrangement of the understorey vegetation, these are only one component of the fuel complex (Gould et al 2011, Cruz et al 2022).In formulating their hypothesis Zylstra et al (2022a) neglect the contribution of more influential fuel characteristics and type (McCaw et al 2012).In particular, the quantity, availability and structure of fine fuels, the main source of energy released by a fire and primary driver of fuel-driven rate of spread, does not decrease in long unburned stands, but rather maintains a pseudo steady state (Walker 1981, Zazali et al 2021, Tangney et al 2022, Burrows et al 2023).Jarrah (Eucalyptus marginata) trees dominate many SW forests, and their fibrous bark is easily ignited and dislodged from the trunk, contributing to overstorey fuel combustion (i.e.onset of crown fire activity) and profuse spot fire ignitions, leading to rapid increases in forward rate of spread and burned area growth (Luke andMcArthur 1978, Storey et al 2020).The influence of bark fuels on fire spread exceeds that of the understorey vegetation (Gould et al 2007) that form the basis of the self-thinning hypothesis put forward by Zylstra et al (2022a).Bark is charred in fire and its hazard increases with TSF: the reduction of bark fuel and thus spot fire potential can only be achieved through its consumption during prescribed burns or wildfires (Gill et al 1986, Gould et al 2007).

Fire severity is not higher in prescribed burning than in wildfire
Zylstra et al (2022a) also suggest that fire severity may be higher in prescribed burns than in wildfires, so that 'a regime dominated by wildfires may have a less pronounced peak in flammability' .However, local (Dixon et al 2022) and eastern Australian (Price et al 2022) studies confirm that wildfires regularly result in more severe fire relative to prescribed burns.Dixon et al (2022) found wildfires in northern Jarrah forests of Western Australia resulted in >34% more area experiencing canopy scorch or canopy consumption relative to prescribed burns.Price et al (2022) found wildfires had significantly higher severity than prescribed fires and 100% burn cover, compared to 30%-40% area remaining unburnt following prescribed burns.Importantly, Collins et al (2021) showed that prior fire severity could significantly influence whether a subsequent fire left areas unburnt or produced crown consumption, depending on forest type and fire weather conditions.Studies also show prescribed fires have a lower impact on ecosystem components than wildfires (Densmore and Clingan 2019, Cova et al 2023).

Further exploration of ecological context is required
Zylstra et al (2022a) restricted their study area to National Parks, aiming to exclude areas historically subject to recent logging and mining, but, except for slope, did not consider any of the substantial bioclimatic gradients and factors they analysed.In the south of their study area, tall open forests occur on fertile soils receiving >1200 mm of rainfall with onshore flows from the Southern Ocean delivering rain periodically through the summer period.In the north-west, low open forests occur on infertile soils receiving <600 mm of rainfall and experience long, dry summer periods (Andrys et al 2017, O'Donnell et al 2021).Driven by these changes, the structure, vegetation and fuel dynamics, and fire regimes of these ecosystems vary substantially (McCaw et al 2002, Gould et al 2007, Tangney et al 2022).Given the reliance of the authors on ecological mechanisms such as succession, ignoring such significant differences in vegetation and climate further call their study design and inference into question.

What can we conclude about wildfire risk patterns, and how can we learn more?
In this letter we discuss how attributes of the data structure and the analytical approach taken in the analysis of fire history data by Zylstra et al (2022a) interact with low overall fire occurrence, to create an artefact that guarantees the finding of decreasing wildfire likelihood with TSF.Wildfire detectability and data sensitivity analyses suggest that the data analysed cannot be interpreted reliably using these approaches beyond around ∼35 years TSF.An appropriate conclusion from the analysis of these data is that forests burnt <6-7 years previously are indeed substantially (2-20 times) less likely to burn again than those unburnt for longer periods of time, with fire likelihood not significantly changing up to at least 35 years.Beyond TSF ∼ 35, data cannot be interpreted reliably using the techniques applied by Zylstra et al (2022a).
An approach to refining this analysis, and potentially extending the range of TSF it addresses could include: repeating the analysis using contemporary data that add five further years of data from the modern era and correct known large errors; update fire history mapping to improve reliability of TSF data and fire boundaries; incorporate fire severity information in analyses; expand areas of coverage to all forests (not restricted to national parks), but do incorporate a much wider range of covariates and factors-including vegetation type, past logging history, rainfall zone, proximity to rivers, towns, roads, forest edges, fire management zones, etc. Contemporary analytical approaches that go beyond paired t-tests could potentially achieve greater statistical power, as would methods that examine fire outcomes from actual ignitions and fire incursions into fuels of different TSF.
We note that analyses of flammability using the same approach have been made for four different forest types in south-eastern Australia (Zylstra 2018).As the forest data reported for these have even lower mean fire likelihoods and sparser samples (Zylstra 2018), we feel it is likely that conclusions of declining wildfire risk in those forests using this technique are also not valid.
Zylstra et al (2022a) address an important issue: fire management for biodiversity, ecosystem services and wildfire hazard reduction is a significant activity in Australian eucalypt forests.Contrary to their findings, we conclude that all current evidence points to flammability in southwestern Australian forests increasing or remaining unchanged with increasing TSF (beyond ∼8 years), not decreasing.This evidence includes detailed analysis of fire history data and understanding of fuel dynamics and fire behaviour models.Based on the current evidence, we conclude that the idea of allowing forests to 'mature' as a policy to achieve reduced wildfire risk is incorrect, and, to a point, hazardous, potentially leading to mismanagement with long term negative impacts on society and the environment.

Figure 1 .
Figure 1.(A) Sample size declining with time since fire (TSF) for forest area, wildfire area and numbers of years sampled (note log scale).(B) Mean (± SE; grey bars) 'flammability' (as defined by Zylstra et al (2022a)-for each TSF; f = average of proportion of forest area that experienced wildfire in each year sampled).Dashed lines indicate mean flammability for all TSF (f mean (all) = 2.04%), 'young' (TSF 1-7 years; f mean (young) = 1.01%) and 'mature' forests (TSF 57-64 years; f mean (mature) = 0.31%).Open diamond indicates f1 flammability corrected for single large error (see text).Vertical blue lines indicate mean sensitivity of f in each TSF to fire area varying by ±1 ha in each sample year in turn.(C) The proportion of samples in each TSF that are smaller than 50 ha, the detection limit for mean flammability across all TSF, and; the flammability that would be assessed if precisely 2.04% of all samples experienced fire (corresponding to f mean (all) ), with 1 ha spatial resolution (fire area is recorded at integer ha precision).(D) Change in projected future wildfire extent relative to 2018, based on the f of each TSF (figure (B)), and the distribution of TSF in 2018.f is not reported for TSF >64 years, so TSF >64 was assigned f equal to the average f of either all mature TSF, or all TSF.An additional projection is included (pink line) demonstrating data sensitivity to tiny errors among 'mature' TSF samples-adding one single hectare of wildfire in one sample year (2017) and TSF (62) substantially impacts outcomes-and hence reliability of f estimates for longer TSF.The f1 error shown in figure (B) was corrected in projections by assuming 0 fire in TSF = 1 forests for 2003.
Miller et al Hamilton T, Wittkuhn R and Carpenter C 2009 Creation of a fire history database for southwestern Australia: giving old maps new life in a geographic information system Conservation Sci.Western Aust.7 429-50 Luke R H and McArthur A G 1978 Bushfires in Australia (Australian Government Publishing Service for CSIRO) McCaw L W, Gould J S, Phillip Cheney N, Ellis P F M and Anderson W R 2012 Changes in behaviour of fire in dry eucalypt forest as fuel increases with age For.Ecol.Manage.271 170-81 McCaw W, Neal J and Smith R 2002 Stand characteristics and fuel accumulation in a sequence of even-aged Karri (Eucalyptus diversicolor) stands in south-west Western Australia For. Ecol.Manage.158 263-71 O'Donnell A J, McCaw W L, Cook E R and Grierson P F 2021 Megadroughts and pluvials in southwest Australia: 1350-2017 CE Clim.Dyn.57 1817-31 Plucinski M, McCaw W, Gould J and Wotton B 2014 Predicting the number of daily human-caused bushfires to assist suppression planning in south-west Western Australia Int.J. Wildland Fire 23 520-31 Price O H, Nolan R H and Samson S A 2022 Fuel consumption rates in resprouting eucalypt forest during hazard reduction burns, cultural burns and wildfires For.Ecol.Manage.505 119894 Storey M A, Price O F, Bradstock R A and Sharples J J 2020 Analysis of variation in distance, number, and distribution of spotting in southeast Australian wildfires Fire 3 10 Tangney R, Miller R G, Fontaine J B, Veber W P, Ruthrof K X and Miller B P 2022 Vegetation structure and fuel dynamics in fire-prone, Mediterranean-type Banksia woodlands For.Ecol.Manage.505 119891 Walker J 1981 Fuel dynamics in Australian vegetation Fire and the Australian Biota ed A M Gill et al (Australian Academy of Science) pp 101-27 Zazali H H, Towers I N and Sharples J J 2021 A critical review of fuel accumulation models used in Australian fire management Int.J. Wildland Fire 30 42-56 Zylstra P J 2018 Flammability dynamics in the Australian Alps Austral Ecol.43 578-91 Zylstra P J, Bradshaw S D and Lindenmayer D B 2022a Self-thinning forest understoreys reduce wildfire risk, even in a warming climate Environ.Res.Lett.17 044022 Zylstra P J, Lindenmayer D B and Bradshaw S D 2022b Coming of age: research shows old forests are 3 times less flammable than those just burned The Conversation (available at: https://theconversation.com/coming-of-age-researchshows-old-forests-are-3-times-less-flammable-than-thosejust-burned-179571) (Accessed 23 March) B P