Introduction

Both fire and invasive vegetation are major, non-independent forces shaping vegetation composition and structure in naturally fire-prone tropical savannahs (D’Antonio and Vitousek 1992; Foxcroft et al. 2010; Lindsay Cunningham 2012; Alba et al. 2015). Land managers use fire to reduce fuel build-up, which is thought to reduce the impact of wildfires on flora, fauna, and the built environment (Queensland’s Fire and Rescue Authority Act 1990; Price et al. 2012). Invasive vegetation can increase the fuel load, thereby incurring more severe environmental effects (Vogler and Owen 2008; Setterfield et al. 2010; Russel-Smith et al. 2012). More frequent fires open the understory vegetation structure in savannahs and open woodlands (Burgess et al. 2014; Alba et al. 2015), potentially negatively impacting fauna that shelter in grasses (Barlow and Peres 2004; Smith et al. 2013), and reducing the availability of shelter sites such as logs, hollows, and tree trunks (Setterfield et al. 2010; Haslem et al. 2011; Russel-Smith et al. 2012; Tng et al. 2014).

Changing fire regimes, in association with changing land use and encroachment of invasive vegetation, are suspected of contributing to declines in small- and medium weight-range mammals in Australia (Johnson 2006; Griffiths and Brook 2014), but few studies have examined the response to fire of mammal communities in tropical savannahs, in Australia, or elsewhere. Different species may display different responses to fire because of differences in resource requirements, or susceptibility to predators, or other ecological or demographic factors. Some mammals are sensitive to frequent fires (Pardon et al. 2003; Andersen et al. 2005; Francl and Small 2013; Griffiths and Brook 2014; Kelly et al. 2014; Griffiths et al. 2015; Mendonça et al. 2015) and gradually decline with repeated burning, or suddenly collapse in abundance with a slow recovery after fire (Pardon et al. 2003; Griffiths and Brook 2014; Griffiths et al. 2015; Mendonça et al. 2015). Some mammals avoid burnt habitats by moving into unburnt sites (Clarke and Kaufman 1990), and therefore occur in lower abundances directly after fire, but return to pre-fire abundances with emerging vegetation cover (Vieira 1999; Kirchner et al. 2011). On the other hand, some mammal species prefer burnt habitats and increase in abundance following fire (Vieira 1999; Breed and Ford 2007; Bock et al. 2011). Overall, there is a wide range of possible responses to fire of different mammal species.

Here we investigate mammal responses to fire in an Australian native savannah invaded by invasive grader grass (Themeda quadrivalvis). We compared replicate sites; each dominated either by native kangaroo grass, native black spear grass, or by non-native, invasive grader grass. We quantified the short-term influence of fire on mammal communities in these sites after they had not been burnt for 2 years, directly after burning, and more than a year after burning when grass cover had returned to pre-fire levels. Our study aimed to provide new insight on the response of these mammals to fire management in naturally fire-prone environments, i.e., open woodlands that often burn more than once per year.

Methods

Study system, sampling sites and periods

We conducted our study in the savannah and open woodland at Undara volcanic National Park in north Queensland, Australia (18°19′29.92″S, 144°36′28.31″E). In total we used 24 sampling sites with 8 replicates of each grass type, such that there were 8 sites each in native kangaroo grass (Themeda triandra), native black spear grass (Heteropogon contortus), and non-native invasive grader grass (Themeda quadrivalvis). In the first year we established 12 sampling sites (4 in each grass type), and each site was sampled for 21 days, we extended our sample to 24 sampling sites (by adding an additional 12 sites) in the second year and trapped each site for 11 days. We selected 50 × 50 m sampling sites, and ensured that there was no spatial clumping of particular grass types. This was possible because of the highly heterogeneous nature of the grasses growing in that area, and because interspersed patches (of a minimum 50 × 50 m) of all three grasses were widely represented at Undara (Supporting information Appendix S1). We targeted these large patches to investigate the mammal community in the different dominant grasses as sites transitioned through three different stages: sites not burnt for 2 years (pre-burnt), directly after burning (post-burnt), and when grass cover had returned to pre-fire levels (revegetated). We trapped small- and medium-sized mammals (5 g–3.5 kg) over eight trapping periods for 2 years in four distinct trapping periods per year: in the pre-wet (21 Oct–14 Nov 2008 and 2009), mid-wet (3–26 March 2009 and 2010), early-dry (14 April–6 May 2009 and 2010), and mid-dry seasons (14 July–12 Aug 2009 and 2010) with number of sites in pre-burnt, post-burnt, and revegetated state per trapping period (Table 1).

Table 1 Untransformed mammal abundance, richness, and abundance of individual mammal species
Full size table

Site history, grasses, and fire

All of our sampling sites were located in savannah open woodland, and more detailed information about the area and sites including climatic data and in-grass temperatures for our study period can be found in Abom et al. (2015). In addition to the woodland, depressed lava tubes meander through the landscape, characterised by evergreen vine thicket vegetation similar to that found along the east coast of tropical Australia (Atkinson and Atkinson 1995), although we did not sample these areas. Prior to 1992, when Undara became a national park, the area was a grazing property, and some parts were used for growing vegetables. Remnants of the fields can still be seen to the east of Yarramulla Ranger Station (at sites 2G, 7G, and 8G, Supporting information Appendix S1).

Grader grass is a common invasive grass in disturbed systems worldwide. It is an annual grass, native to India, and is considered a major threat to natural, cultivated, and recreational areas where it is introduced (Keir and Vogler 2006). Grader grass has spread rapidly throughout central and northern Australia. It is a typical invasive grass, it emerges as a single stolon (up to 3 m tall), is fast growing, and a prolific seeder that can germinate all year round in northern Australia (for a more comprehensive review see Keir and Vogler 2006). In contrast, native kangaroo (Themeda triandra) and native black spear (Heteropogon contortus) grasses are perennials and grow in clumps or hummocks to 1.5 m in height.

Land managers (including park rangers at Undara) use fire to reduce invasive grader grass cover and lessen the impact of wildfires in the hotter pre-wet season (October–December), and burn selected areas at Undara on rotation in the cooler early dry season (April–May) to create a mosaic of burnt and unburnt areas (30–60 % burnt, the rest unburnt, Queensland’s Fire and Rescue Authority Act 1990). Savannah woodlands are naturally fire-prone (Foxcroft et al. 2010), and Undara Volcanic National Park is burnt at least partially by wildfires every 3–5 years. In the current study, sampling sites had been burnt on rotation every 2 years since 2002, with wildfires in October 2003 that burnt the entire park and in November 2008 that burnt areas of the park (Appendix S1). During the current study, park rangers applied prescribed fires in April of 2008, 2009, and 2010 when environmental conditions were cool enough to allow the fire to self-extinguish in the late afternoons (Appendix S1; Queensland’s Fire and Rescue Authority Act 1990). However, some prescribed fires escaped causing uneven numbers of sampling sites. Thus, some sites (three grader grass, and two spear grass) were only sampled as post-burnt and revegetated sites and not in the initial pre-burnt stage, and two spear grass sites were sampled only as unburned and burned sites, and not after revegetation.

Habitat and mammal sampling protocol

To track changes in habitat variables and vegetation cover in pre-burnt, post-burnt, and at revegetated sites, we sampled each site in each trapping period by laying four 50-m transects transversely, spaced evenly 16.6 m apart, and recording each variable in linear centimetres along the 2.5 cm-wide transect tape. The data collected from the four transects were combined to create a mean  % cover of each variable per site. We recorded % cover of grader, kangaroo, and black spear grass (these were the dominant grasses), mixed grass (% cover of all other grasses combined), broad-leaved vegetation (herbaceous, legumes, and small bushes), leaf litter, logs, rocks, bare ground (i.e., ground without vegetation, rocks or leaf litter), burnt areas blackened by fire and tree canopy cover above the transect (trees up to 18 m tall, but more commonly between 8 and 12 m tall).

To describe mammal composition at the sampling sites, we used pitfall traps, and baited Elliott and cage traps (Appendix S1). In the centre of each 50 × 50 m sampling site, 5 unbaited pitfall (20-l straight-sided bucket) traps were deployed, in an X-shape with one centre pitfall trap and four equally spaced arms. Traps were spaced 10 m apart and connected via a 0.5 m high drift fence (Cyclone MeshTM). We lined each pitfall trap with a 5-cm layer of leaf litter and provided a moistened sponge for water and cover to captured animals. At each trap site we used 12 baited Elliott traps (W 100 × L 325 × H 100 mm) spaced 10-m apart encompassing the pitfall trap array, and at the outer perimeter of the trapping area, we placed four baited cage traps (W 300 × L 605 × H 290 mm) in a square, spaced 50-m apart. Elliott and cage traps were placed in a naturally shaded area or a shade cloth was provided. We used a mixture of oats, vanilla essence, and peanut butter as bait in Elliot and cage traps. We baited Elliott and cage traps every second day in the early evening (17:00–19:00) and checked, cleared, and closed Elliott and cage traps in the early morning (04:30–06:30). Pitfalls remained open 24 h a day and we monitored these traps twice daily, in the early morning (05:00–08:00), and in the late afternoon (16:30–18:30). Prior to release of mammals at their point of capture, mammals were identified to species level using Menkhorst and Knight (2004). We marked medium-sized mammals individually using ear clips, and we batch-marked small mammals by trimming the tip of the tail with a pair of scissors (Livingstone, SDI) to obtain DNA samples for another study, and to distinguish between captures and recaptures. Clipping tools were sterilized with an open flame between individuals.

Statistical analyses

Habitat and mammal assemblage analyses

We tracked and described the changes in vegetation cover in pre-burnt, post-burnt, and revegetated sampling sites, which provided information on the progress of each dominant grass type (grader, kangaroo, and black spear grass). Environmental variables (mean  % cover of each dominant grass, and mixed grass, leaf litter, logs, rock, bare ground, burnt area, and canopy) at each sampling site were first averaged over the four transects per site, and then site variables were averaged across the number of trapping periods for which the sampling site remained in one of the three states (pre-burnt, post-burnt, or revegetated). Mammal trapping data were standardised to 100 trap nights. Prior to statistical analyses, all environmental variables and mammal abundance and richness were standardised using a relativising transformation to a proportion between 0 and 1, by dividing each site variable, mammal abundance, and richness, by the maximum number of that variable at any sampling site. This procedure helped prevent very abundant variables driving the results.

To investigate patterns in site variables among sampling sites and states, we used the statistical package PC-ORD (McCune and Mefford 1999). We used non-metric multidimensional scaling (NMDS) to explore patterns among sites in the quantitative variables: mean % cover of grader, kangaroo, black spear, and mixed grass, leaf litter, logs, rock, bare ground, burnt area, and canopy cover (cut-off value r2 ≤ 0.20) in the three different categorical site states: pre-burnt, post-burnt, and revegetated grass sites. For the NMDS analyses, we used the autopilot “slow and thorough” with Sorensen distance measures, dimensionality was determined by Monte Carlo tests (9999 permutations, significance test of stress in relation to dimensionality of the number of axes in the final analysis). We extracted axis and cumulative scores using a Bray-Curtis (Sorensen) dissimilarity index with original end point selection, city-block projection geometry and calculation of residuals. To illustrate site trends among treatments, we constructed bi-plots from NMDS sites and environmental variable scores (McCune and Mefford 1999).

To investigate the effect of our treatments, i.e. fire (pre-burnt, post-burnt, and revegetated) and dominant grass species (grader, kangaroo, and black spear grass), and their interactions, on mammal abundances and richness, we used generalised linear mixed effects models (for each respective mammal species’ relativized abundance, overall abundance, and overall richness), with a Gaussian error structure, in the statistical program R v2.15.2, lmer function in the package lme4, (R Development Core Team 2012; Bates et al. 2013). All response variables were checked for normality and transformed accordingly prior to this analysis. Three model setups were used to partition out the effects of fire, grass treatments and their interactions, always maintaining sites as a random factor, i.e. abundance or richness as a function of: grass type (as a fixed effect) + fire treatment (as a random effect) + site (as a random effect); fire treatment (as a fixed effect) + grass type (as a random effect) + site (as a random effect); and grass type * fire treatment (as fixed effects) + site (as a random effect). To determine if treatments differed significantly, we then conducted pairwise comparisons of the fixed effects (all combinations of grass type or fire treatments and their interactions, respectively) using Tukey’s adjusted Least Squares Means, via the “lsmeans” function in the R package “lsmeans” (Lenth 2013).

To then determine the relative importance of the site-specific environmental variables on mammal abundance and richness within each treatment, we again used GLMMs, via the lmer function in the lme4 package (Bates et al. 2013). Followed by a dredge function (automated model selection) and model averaging in the MuMIn package (Barton 2013) in R. In these models, fire treatment and site were input as random factors, with grass type included in the fixed effect variables with the other % cover site variables. Thus, we first made a global model for each abundance and richness measure, inputting all 11 environmental variables (i.e., grader, kangaroo and spear grass cover, bare ground, broad leaf, burn area, canopy, leaf litter, log, rock and mixed grass cover) as fixed effects, and then we used the dredge function to rank all possible models (Barton 2013). In this approach, we limited the fixed effects to one variable per model (i.e. setting m.max in dredge to = 1), as prior analysis showed that models with >1 environmental variable (fixed effect) were often over-fitted or weaker than simpler models. Then, we used model averaging with multimodel inference to investigate the relative importance of each environmental variable to mammals (Barton 2013). Models were ranked according to model fit using the corrected Akaike information criterion (AICc), and models with ≤2 ΔAICc were considered the most parsimonious and descriptive of the data (Burnham and Anderson 2002). Once the best explaining variables/models were defined, we tested all top models (≤2 ΔAICc, i.e. with one fixed effect) for over-dispersion, by dividing the Pearson residuals by the residual degrees of freedom of the model, and determining probability with the χ2 distribution. No models were over-dispersed (χ2 P ≥ 0.05).

Results

Habitat composition

We found a stable two-dimensional NMDS solution accounting for 90.43 % of the variance (first axis = 58.43 %, and second axis = 32.00 %) with a final stress of 0.086 for sites and environmental site variables (Fig. 1a, b), and mean cover % ± SE of habitat variables (Table 1). Sampling sites showed strong patterns in vegetation cover among dominant grass sites and states: pre-burnt, post-burnt, and revegetated grass sites clustered into separate groups (Fig. 1a). Pre-burnt and revegetated grader grass grouped away from native grass sites, and post-burnt grass sites clustered in three distinct groups away from pre-burnt and revegetated sites (Fig. 1a). Interestingly, all pre-burnt grass sites had very similar dominant grass cover, but one large difference among grass types after burning was that grader grass grew back more uniformly. This trend towards more uniform revegetation was not, however, detected in native grass sites (Table 1, and Supplementary information Appendix S2). The mixed grass cover was more strongly associated with pre-burnt states, while broad-leaved vegetation was more associated with revegetated grader grass sites. Bare ground, leaf litter, log, and canopy cover were more closely associated with pre-burnt and revegetated native grass sites. The percent cover of burnt area, as expected, was strongly associated with burnt grass sites (Fig. 1b).

Fig. 1
figure 1

a Vegetation structure as a two-dimensional NMDS ordination (stress = 0.086). The first axis represents 58.43 % of the variation, and the second axis 32.00 %. Symbols; pre-burnt (open), post-burnt (filled), and revegetated (red) with grader = circles, kangaroo = triangles, and black spear grass = squares. b Environmental variables driving the NMDS results (r2 > 0.20)

Full size image

Mammal captures

We sampled for a total of 24,960 trap nights, and captured a total of 1029 mammals (467 individuals, excluding recaptures) from 10 different species (untransformed catch numbers are provided in Table 1). Excluding recaptures, the eastern chestnut mouse (Pseudomys gracilicaudatus, n = 137), and northern brown bandicoot (Isoodon macrourus, n = 124) were the most common mammals, followed by rufous bettongs (Aepyprymnus rufescens, n = 89), house mice (Mus musculus, n = 47), common planigales (Planigale maculata, n = 35), tropical short-tailed mice (Leggadina lakedownensis, n = 24), brush tail possums (Trichosurus vulpecula, n = 6), feral cats (Felis catus, n = 2), rabbits (Oryctolagus cuniculus, n = 2), and a stripe-faced dunnart (Sminthopis macroura, n = 1).

Movement among sites

Recaptures showed that individually marked northern brown bandicoots and rufous bettongs moved greater distances than small mammals, and moved among sites. Hence, we used recaptures of these animals to examine habitat use. 49 % of marked northern brown bandicoots (n = 181 of 371 recaptures) and 59 % of marked rufous bettongs (n = 113 of 193 recaptures) were recaptured at sites within 5 km of their previous capture site. Northern brown bandicoots were recaptured most often in pre-burnt native kangaroo (n = 53, 29 %) and black spear grass (n = 50, 28 %), and we detected bandicoots using pre-burnt grader grass only 10 % (n = 18) of the time. Northern brown bandicoots had the lowest recaptures in post-burnt and revegetated grass sites (≤ 8 %).

Rufous bettongs were recaptured more frequently in pre-burnt black spear grass sites (n = 37, 33 %) with only 6 % or less recaptured in pre-burnt grader and kangaroo grass sites. However, once sites had been burnt, we found that rufous bettongs were recaptured more frequently in post-burnt (n = 14, 12 %) and revegetated kangaroo grass (n = 18, 16 %) than in similar states in grader and black spear grass (≤9 %). In contrast, the small mammals, including eastern chestnut mice, house mice, tropical short-tailed mice, and common planigales were not recaptured at sites other than their site of initial capture, verified using tail clips with data on size and sex of marked individuals at each site.

Patterns of change in overall richness, abundance and individual species abundance in response to fire in different grasses.

Pairwise comparisons among richnesses and abundances in dominant grass types, using Tukey’s least square means from the GLMM analysis, did not show any significant effect for any combination of grass type (P ≥ 0.05 for all combinations, results not shown). We detected differences in abundance and richness among fire treatments (Fig. 2 and Appendix S3 for full result table). Mammal richness was relatively low, and there were no significant differences in mammal richnesses among the different fire states (pre-burnt, post-burnt, and revegetated). Mammal abundance, however, was significantly greater in pre-burnt than in post-burnt sites (Tukey’s P = 0.003), and remained lower, even when grass cover had returned to the same level as it was in preburnt sites, once revegetated (P ≤ 0.001). Overall, mammal abundances were between 40 and 55 % lower in revegetated sites compared to pre-burnt sites (Fig. 2).

Fig. 2
figure 2

Mean mammal abundances versus burn treatment (pre, post and revegetated). Abundances are untransformed and standardised to 100 trap nights in grader (dark grey), kangaroo (light grey), and black spear (black) grass (±SE; zero values = no animals captured). Treatments with differences in relativized abundances (see text for variable definition), from the generalized linear mixed effects models, showing effects of fire while controlling for site and grass type, are denoted by having different letters (a–b) shown above the bar/grouping. Level of effect for differing treatments are represented with P, i.e. P < 0.05 *; P < 0.01 **; P < 0.001 ***

Full size image

There was a significant interaction between responses in grass type and fire only for northern brown bandicoot abundances. Northern brown bandicoots had lower abundances in pre-burnt compared to revegetated kangaroo (t-ratio = −5.30, df = 23, P = 0.001, Fig. 2) and spear grass sites (t-ratio = −4.09, df = 26, P = 0.009), and lower in post burnt compared to pre burnt kangaroo grass (t-ratio = −3.59, df = 23, P = 0.03). No significant interactions between grader grass and fire on abundances were found for any overall abundance or richness measure. Results for other abundance and richness measures for the interactive models are also not shown due to lack of significant effects. These results were supported in that brown bandicoot abundances were similar in all the pre-burnt grass sites, but were significantly higher in pre-burnt grass than in post-burnt (P ≤ 0.001) and revegetated grass (P ≤ 0.001, Fig. 2). In contrast, rufous bettongs increased in abundances in post-burnt, compared to pre-burnt, sites (P = 0.033, Fig. 2).

When controlling for grass type and site (fire treatment as fixed effect), models showed that eastern chestnut mice were more abundant in pre-burnt than in post-burnt (P = 0.018) sites, while house mice were detected in higher abundances in pre-burnt than in post-burnt (P = 0.008) and revegetated sites (P = 0.002, Fig. 2). Interestingly, eastern chestnut mice were most abundant in grader grass before burning, and reached their lowest abundances in all post-burnt grass sites (Fig. 2). Abundance of eastern chestnut mice returned to pre-burnt levels in grader grass once the grass had regrown, but did not return to prior abundances in revegetated native grasses (Fig. 2). In contrast, house mice abundance did not recover with emerging grass cover (Fig. 2). Abundances of the tropical short-tailed mouse were higher in post-burnt grasses than in revegetated grasses (P = 0.039, Fig. 2), while common planigales were more common in revegetated grasses than in post-burnt grass sites (P = 0.052, Fig. 2).

Possible reasons for responses to fire: habitat features influencing mammal abundance and richness.

The model selection indicated that critical habitat variables defining mammal distributions varied among mammal species. For some response variables, individual habitat characters held less explanatory power, with the null models of fire treatment and site (random effects) providing the best explanation of the data for all mammal abundance, richness, and rufous bettong abundance (ω i  = 0.26, 0.99, 0.30 respectively, Table 2). Full models for variable selection for the tropical short tailed mouse were over-fitted with the addition of site as random effect (high site dependence in distribution); thus, only the random effect of treatment contributed towards an accurate GLMM for this species. This null model (a site effect) provided the best explanation of variation in tropical short-tailed mouse abundances (Table 2).

Table 2 Model output examining habitat features influencing overall mammal abundance, richness, and individual species abundances, with treatment (pre-burnt, post-burnt, and revegetated sites) and site included as random effects
Full size table

Leaf litter cover was the most supported environmental variable explaining mammal abundance (ω i  = 0.25), richness (ω i  = 0.99) and northern brown bandicoot abundance (ω i  = 0.33, Table 2), aligned with increased abundance values (Fig. 3). Bare ground also lead to high abundances of all mammals together, and bettongs and bandicoots alone (Fig. 3; Table 2).

Fig. 3
figure 3

Regression coefficients with 95 % confidence intervals from the generalised linear mixed effects models explaining mammal abundance in pre-burnt, post-burnt and re-vegetated sites. Variables with the greatest influence (ΔAICC ≤ 2, see text) are denoted with an asterisk

Full size image

For small mammals, broad-leaved vegetation was strongly and positively related to abundances of common planigales (ω i  = 0.98), eastern chestnut mice (ω i  = 0.54) and house mice (ω i  = 0.26) (Fig. 3; 2). Tropical short-tailed mice were more abundant with greater rock cover (w i  = 0.24) and leaf litter (ω i  = 0.16, Table 2; Fig. 3), and with burn area (a less parsimonious model but significant correlation, Fig. 3).

Discussion

Mammal species richness remained similar in pre-burnt, post-burnt, and revegetated grass sites, whereas mammal abundance was more variable among grass types and states. We captured more mammals in pre-burnt than in post-burnt sites, and the lowest abundance of mammals was recorded in revegetated sites 15 months after burning. The interaction between fire and grader grass seemed to homogenize the habitat, as grader grass grew back as a more complete monoculture, and fire in grader grass sites removed more habitat features such as logs and leaf litter, compared to native grass (Abom and Schwarzkopf 2016). These observations are all consistent with previous studies, but our focus on individual species showed that the habitat changes wrought by fire seemed to discourage only some mammal species, whereas others were equally (or even more) abundant after fire and grass regeneration.

For example, bandicoots declined in our study after fire. Similarly, in another study (the Kapalga fire experiment), a bandicoot (Isoodon macrourus) population declined with increased fire frequency from about three animals per 100 trap nights to one bandicoot in 7000 trap nights, and even after five years of fire exclusion, the bandicoot population had not recovered (Pardon et al. 2003). In contrast, we observed increases in tropical short-tailed mice in response to fire, similar to various small mammals in other experiments involving fire (Calomys callosus in Brazil, Vieira 1999; Peromyscus maniculatus in Kansas, Clark and Kaufman 1990; and Chaetodipus hispidus, C. baileyi, Perognathus flavus in Arizona, Bock et al. 2011). In our study, eastern chestnut mice returned with returning grass, and similar population cycles have been described in other small rodent populations (e.g., Rattus villosissimus in central Queensland, D’Souza et al. 2013; and Bolomys lasiurus in the Brazillian Cerrado savannah, Vieira 1999).

How did fire influence habitat?

Pre-burnt, post-burnt, and revegetated dominant grass sites were distinguishable in terms of environmental variables. Mixed grass cover was strongly associated with pre-burnt grader grass sites, while broad-leaved vegetation was more strongly associated with revegetated grader grass sites. In contrast, percent cover of bare ground, leaf litter, logs and canopy were all higher in pre-burnt and revegetated kangaroo and black spear grass sites. Rocks were more visible in post-burnt areas, and grouped with burnt grass sites. Interestingly, the vegetation composition of pre-burnt grader grass was more variable than revegetated grader grass sites, indicating that grader grass grew back as a purer stand, providing evidence for the homogenizing effects of fire when applied to invasive grasses (Table 1, Supplementary information: Appendix S2; Abom and Schwarzkopf 2016). In contrast, the site composition in native grass changed little following fire, indicated by the grouping of pre-burnt and revegetated native kangaroo and black spear grass sites. Other studies investigating invasive grass growth after fire have also found that invasive grasses grow back more densely than native grasses, reducing plant diversity and habitat heterogeneity in native systems invaded by non-native vegetation (Vogler and Owen 2008; Lindsay and Cunningham 2012; Alba et al. 2015; Abom and Schwarzkopf 2016). The clearing of grasses (native and invasive) by fire promotes the establishment of invasive vegetation, which has positive effects on its proliferation (Foxcroft et al. 2010; Setterfield et al. 2010; Alba et al. 2015; Abom and Schwarzkopf 2016). Invasive grasses with higher dead standing biomass burn hotter than native grasses, and these hotter fires simplify the savannah by consuming low understory vegetation, and fallen logs which provide structure and hollows used by many animals (Setterfield et al. 2010; Haslem et al. 2011; Russel-Smith et al. 2012; Tng et al. 2014; Abom and Schwarzkopf 2016).

How did fire and invasive grader grass influence mammals?

A drawback of the current study was a lack of control sites, and the high frequency of prescribed burns, which did not allow us to examine differences among habitats affected by different fire frequencies. In spite of these issues, we were able to distinguish patterns in response to fire. Mammals occurred in higher abundances in pre-burnt grass sites, which was interesting because our sampling sites have been burnt every two years since 2002, and at times more often, when there have been wildfires (e.g. 2003). We did not, however, find that fire in invasive grader grass was worse for mammal abundances than fire in native grasses, instead, the response in terms of overall abundance was simply lower in all grasses after fire. An inability to detect a significant interaction in overall mammal abundance between invasive grass and fire, even though there were clearly differences among mammal responses, suggests that mammals were not responding especially strongly to the reduced habitat features we detected in grader grass. In addition, some species increased in abundance following fire in grader grass. In most grasses, bettongs were positively influenced by fire, perhaps because fire allowed easier access to buried food, such as truffles (Vernes and Pope 2001; Pope et al. 2005). Short-tailed tropical mice also increased after fire, although the reason for this is unclear. They may move more in habitats created by fire than they do in undisturbed habitats (Moro and Morris 2000; Kutt and Kemp 2005), or prefer open habitats and the exposed rocks associated with them. These mice were detected in reduced numbers after vegetation regrew.

Medium-sized mammal species, like the northern brown bandicoot, had a strong preference for pre-burnt native grasses and were strongly linked to native grass sites characterised by a moderate amount of bare ground and leaf litter. Bandicoot numbers declined in post-burnt areas, and their abundances declined further in revegetated grasses, possibly due to the reduction in cover of leaf litter and thus, habitat characteristics associated with high leaf litter (e.g., such as specific vegetation types, soil organisms and invertebrate populations). Other studies have shown that the northern brown bandicoot is sensitive to frequent and large scale fires (Pardon et al. 2003; Woinarski et al. 2004; Griffiths et al. 2015). Reduced numbers of bandicoots may be caused by high fire-induced mortality at the time of fire, higher post-fire deaths, or emigration to more suitable habitats (Pardon et al. 2003; Friend 1990). In our study, the prescribed fires were ignited in the early dry season (April 2008 and 2009) and a wildfire occurred in the late dry season (October 2008), which burned parts of Undara, but not our sampling sites. Our northern brown bandicoot population behaved very similarly to those described previously (Friend 1990; Pardon et al. 2003). Bandicoots were initially abundant (~5 individuals per 100 trap nights, Fig. 2) in pre-burnt grass habitats, and had ~50 % lower capture rates in post-burnt grasses, decreasing to less than one individual per 200 trap nights in sites in vegetation that had regrown after fire. Like other northern brown bandicoot populations, bandicoots at Undara did not collapse rapidly; instead the population declined gradually over 15 months (after May 2009 to July 2010) until bandicoot numbers were very low.

Small mammals showed mixed responses to fire. Common planigales, eastern chestnut mice and house mice were detected in higher abundances in pre-burnt grass sites with more broad-leaved vegetation. Planigales and chestnut mice were positively associated with grader grass cover. Although planigale, chestnut and house mice numbers were much reduced in post-burnt grass sites, only planigales and chestnut mice were negatively associated with burnt area, and returned with emerging grass in revegetated sites. House mice, on the other hand, almost disappeared once sites had been burnt.

Conclusions

Tropical savannahs are highly diverse and naturally fire-prone systems (Foxcroft et al. 2010). We found, as have other studies, that fire in invasive vegetation changes the vegetation structure, reduces plant biodiversity (Setterfield et al. 2010; Haslem et al. 2011; Russel-Smith et al. 2012; Burgess et al. 2014; Griffiths et al. 2015), and reduced overall mammal abundance to an extent similar to the reduction in native grasses. Land managers use prescribed fires with the intention of reducing invasive vegetation, and lessening the impact of wildfires on wildlife (Price et al. 2012). This approach may not be successful for fire sensitive mammals (Griffiths and Brook 2014). For example, our observed decline in northern brown bandicoots was most likely because of reductions in suitable habitats. For this species, frequent fire may increase the risk of local extinction (Pardon et al. 2003; Griffiths et al. 2015). On the other hand, some mammals appeared to prefer the habitat created by fire. In our study, rufous bettongs and tropical short-tailed mice were most common immediately after burning, becoming less abundant as vegetation returned. Finally, eastern chestnut mice returned to their previous abundances in invasive grader grass, but not in native grass, while rufous bettongs returned to their previous abundances in native grass, but not in invasive non-native grader grass sites. Here we suggest that prescribed burns in naturally fire-prone systems may reduce overall mammal abundance, and if the conservation goal is to avoid reductions in abundance of mammals generally, we recommend that areas be burnt less often, or less regularly, potentially allowing a more diverse vegetation assemblage and structure to be established (Parr and Anderson 2006; Burgess et al. 2014; Griffiths et al. 2015). On the other hand, we found that the responses of individual species to fire in grader grass were idiosyncratic, and some species may prefer burned areas. Thus, if the conservation goal is to maintain biodiversity, we must identify which species are more sensitive to frequent fires, and establish fire regimes based on multiple mammal species’ response (Litt et al. 2011; Kelly et al. 2014).