Habitat structure facilitates coexistence of native and invasive mesopredators in an Australian tropical savanna

Keywords: conservation, ecology, environment, habitat use, introduced species, invasive species, predator–prey interactions, threatened species, wildlife management.

Invasive predators are a serious threat to biodiversity globally (Blackburn et al. 2004; Doherty et al. 2016; Dueñas et al. 2021). Some of the most important ways predators can affect ecological communities are by killing prey species and suppressing and/or excluding sympatric predators (Tannerfeldt et al. 2002; Cáceres et al. 2009). These impacts can be strongly accentuated when native species exhibit naivety towards invasive predators, typically through a lack of shared evolutionary history (Sih et al. 2010). As a result, 142 terrestrial species have become extinct through the impacts of invasive mammalian predators, with a further ~600 species still being threatened globally (Doherty et al. 2016).

Many invasive mammal species have benefited from human-assisted migration and are now firmly established across the globe (Long 2003; Clout et al. 2008). The introduction of the cat (Felis catus) to Australia by Europeans in the late 18th Century has had a catastrophic impact on native wildlife (Abbott 2008; Doherty et al. 2017). Taking advantage of naïve native prey, the proliferation of introduced prey species, and the paucity of large, native predators, cats quickly dispersed across the continent, with feral populations becoming established in all of Australia’s bioregions (Abbott 2008; Doherty et al. 2014; Legge et al. 2020). The extinction of at least 26 of Australia’s native mammal species has been associated with cat predation, with a further ~75 native mammal species being currently threatened (Woinarski et al. 2019). Invasive predators, such as the cat, are still a significant threat to many native mammal species, posing an urgent need to curtail their impact on threatened species.

Efforts to control feral cat populations are often limited by scale, effort, and cost (Campbell et al. 2011; Parkes et al. 2014; Baker and Bode 2016). Complete eradication of invasive predators is seen as the most beneficial, long-term solution to conserving threatened fauna (Jones et al. 2016; Holmes et al. 2019), but to do so often requires extensive monitoring to ensure success, can be cost prohibitive at large scale and, even when conducted using best practices, can have negative, unintended outcomes (Bodey et al. 2011; Lazenby et al. 2014; Holmes et al. 2019). Successful eradications have occurred on off-shore islands up to 630 km² in area (Campbell et al. 2011) and in increasingly large, predator-free fenced reserves (Legge et al. 2018). However, open landscape-scale control programs are often hampered by the continual immigration of cats into areas with cat control (Doherty et al. 2017). These challenges mean eradication may not always be a viable or appropriate solution to conserving native wildlife threatened by invasive species.

Instead of eradication, manipulation of ecological factors that limit feral cat populations may be a more feasible and cost-effective approach to control cats (Doherty et al. 2015; Miritis et al. 2020). For example, research suggests that in northern Australia’s savanna landscapes cats have a strong association with frequently burnt habitat and open, heavily grazed areas, because the reduced vegetation cover in both cases may improve hunting success (McGregor et al. 2014, 2015; Davies et al. 2020). Australia’s only mammalian apex predator, the dingo (Canis familiaris), has been found to consume cats (Doherty et al. 2019), suggesting a degree of top–down predation pressure on feral cat activity. Although, the role dingoes play in suppressing cat activity is heavily debated (Letnic et al. 2012; Fancourt et al. 2019; Kreplins et al. 2021), with experimental removal of dingoes not having been found to increase feral cat abundances in response (Letnic et al. 2009; Castle et al. 2021). As such, this poses a need to identify how environmental factors limit cat activity and what potential landscape solutions can reduce the impact cat predation has on threatened mammals.

In this study, we investigated the spatial patterns of habitat use by cats and a coexisting native mesopredator, the northern quoll (Dasyurus hallucatus), in tropical savanna at Weipa on Cape York Peninsula, Queensland. The northern quoll is the largest native mammalian mesopredator across northern Australia, and the smallest member of the genus Dasyurus, weighing between ~400 and 1200 g (Van Dyck and Strahan 2008). The current distribution of the northern quoll extends patchily across monsoonal northern Australia (Moore et al. 2022; Fig. 1). Despite the northern quoll’s widespread distribution, it has suffered significant population declines in recent decades (Braithwaite and Griffiths 1994; Moore et al. 2019), and is considered Endangered (Oakwood et al. 2016). One of the major drivers of the decline of the northern quoll, especially in higher-rainfall regions, is poisoning following ingestion of the toxic and invasive cane toad (Rhinella marina; Covacevich and Archer 1975; Burnett 1997). However, other drivers of the decline of the northern quoll are thought to include predation by feral cats, high fire frequencies and habitat degradation by introduced herbivores (Braithwaite and Griffiths 1994; Moore et al. 2022). Lowland savanna populations of northern quolls seem to be the most severely affected by these threats, whereas high-rainfall regions and topographically rugged habitat harbour the largest remnant populations (Braithwaite and Griffiths 1994; Oakwood 2000; Moore et al. 2019). Despite these observed trends, our study population of northern quolls has been found to reside in tropical savanna that lacks extensive topographical ruggedness and is inhabited by both cats and cane toads. As such, this remnant northern quoll population provides an excellent model system to understand how environmental conditions and habitat attributes can contribute to the persistence of native mammals alongside cats and will assist in developing appropriate strategies for managing habitat that facilitates coexistence.

Fig. 1.  Camera-trap grids across (a) the northern plateau and (b) the southern plateau with respect to elevation near Weipa on the Cape York Peninsula, Queensland. White circles represent individual camera-trap sites. Black lines represent the plateau vegetation community. The inset map shows the current and former range of the northern quoll (Moore et al. 2022).

To infer how cats and northern quolls can coexist within tropical savanna, we aimed to identify the environmental variables that were the strongest drivers of cat and northern quoll habitat use. We investigated a range of ground-based and satellite-derived variables that have been previously identified as influencing spatial patterns of cats and native small mammals across the northern Australia, including fire regimes (Woinarski et al. 2010; Griffiths et al. 2015; McGregor et al. 2016; Davies et al. 2020), vegetation structure (Hernandez-Santin et al. 2016; Stobo-Wilson et al. 2020a, 2020b), water proximity (McGregor et al. 2016; Moore et al. 2019) and topographical attributes (Hernandez-Santin et al. 2016; Moore et al. 2019; McDonald et al. 2020). On the basis of these previous findings and the conditions of the plateau environment, we predicted that the environmental conditions and attributes found atop the plateaux would be most frequently used by northern quolls, whereas the environmental conditions and attributes found in low-lying savannas surrounding the plateaux would be most frequently used by cats.

The northern quoll is a highly adaptable, generalist mesopredator that occurs in a variety of habitats, including tropical savanna, monsoon rainforest, and rocky outcrops (Moore et al. 2022). As a generalist omnivore, the northern quoll consumes a varied diet consisting of fruits, seeds, and live prey such as rodents, reptiles and invertebrates (Dunlop et al. 2017). It also uses a wide variety of denning resources such as tree hollows, fallen logs, rock crevices and termite mounds, depending on the availability of hollow resources in the landscape (Oakwood 1997).

The northern quoll was once common across Cape York Peninsula (Fig. 1), inhabiting both the savannas of the western coast and the rainforests of the eastern coast (Braithwaite and Griffiths 1994). It is believed that the spread of the cane toad throughout Cape York Peninsula by the 1980s resulted in a widespread extirpation of northern quoll populations (Burnett 1997). The cane toad was first introduced to Australia in 1935, to control the native cane beetle (Dermolepida albohirtum); however, the cane toad also secrets potent poisons, collectively termed bufotoxin, that is lethal on ingestion by the northern quoll (Ujvari et al. 2013). Despite this prominent threat, the northern quoll was observed to be in decline prior to cane toad arrival and is currently in decline in regions where the cane toad is yet to colonise, such as in the Pilbara region of Western Australia (Braithwaite and Griffiths 1994; Woinarski et al. 2011; Moore et al. 2022). The remnant population of northern quolls on which our study focuses was discovered in 2013, on bauxite plateaux near the town of Weipa (Fig. 1). Prior to that, the northern quoll had not been recorded in the Weipa region since 1985 (Winter and Atherton 1985). The area of occupancy and number of northern quoll populations elsewhere across Cape York Peninsula are not known.

The study area is located on the north-western coast of Cape York Peninsula in north-eastern Australia, east of the town of Weipa. The region has a tropical monsoonal climate, with 90% of the mean annual rainfall of 1700 mm typically falling in the summer wet season (December–April; Australian Bureau of Meteorology 2021). The landscape is characterised by elevated bauxite plateaux, dominated by tall, dense savanna, interspersed between low-lying plains, dominated by open savanna, and river tributaries. The bauxite plateaux absorb large quantities of water during the wet season, which gets slowly released through springheads during the dry season, flowing down channels into the surrounding landscape (Specht et al. 1977; Leblanc et al. 2015). This hydrology allows for the formation of tall savannas with high tree biomass and a dense understorey (Bowman and Fensham 1991).

The study was undertaken across two bauxite plateaux known to be inhabited by both cats and northern quolls. Each plateau is approximately 45 km² in area, with elevations of 20 m and 100 m respectively, above the surrounding low-lying, open savanna. The vegetation of both plateaux consists of a sparse canopy dominated by Eucalyptus tetrodonta 25–35 m in height, with subdominant trees Corymbia nesophila, C. stockeri and Erythrophleum chlorostachys. Fruit-bearing trees such as Pandanus spiralis, Livistona muelleri, Parinari nonda, Xylomelum scottianum and Planchonia careya dominate the mid-storey, and the ground layer is dominated by annual and perennial grasses such as Sarga plumosum, Heteropogon triticeus, Alloteropsis semialata and Eulalia mackinlayi. Large termite mounds and hollow trees, both of which are known to be used for denning by the northern quoll, are common. Each plateau has some rugged, pronounced edges, but most of the plateau boundaries are gentle inclines with extensive ecotonal areas of intergrading vegetation communities. Neither plateau, nor the surrounding habitat, has significant areas of exposed rock.

The plateaux are characterised by reduced fire frequencies compared with the surrounding low-lying savanna (~0.6 fires year⁻¹ for plateaux vs 0.9 fires year⁻¹ for the surrounding low-lying savanna over a 20-year period; North Australian Fire Information 2021). Hazard-reduction burns are undertaken at the northern plateau (Fig. 1a) during the early dry season (May–July), but no such burns occur on the southern plateau (Fig. 1b). High-intensity wildfires occasionally occur, typically in the late dry season (August–November).

Camera grids, targeting northern quolls, were established across both plateaux and immediately adjacent areas in 2018, and surveyed annually to 2020 (i.e. 2018, 2019 and 2020). Each camera grid consisted of 143 motion-activated cameras (PC800 Hyperfire Professional IR; Reconyx Inc., Holmen, WI, USA) at a grid spacing of 500 m × 500 m (Fig. 1). Each camera was attached to a tree with a bracket 1.5 m above the ground. Each camera faced straight down at a bait tube containing dried chicken meat. Cameras were active continuously (i.e. 24 h each day), and programmed to capture three images per trigger, with a high sensitivity. Cameras were deployed for a minimum of 10 nights in April–May of each year, to coincide with the northern quoll breeding season, when individuals are most numerous and active. All 143 cameras deployed within each plateau grid were operated simultaneously. Between plateau grids, there were slight differences in camera operation times due to the time restraints of deployment. In total, 12 599 trap nights were sampled over 3 years, from 286 camera sites across the two plateaux.

Camera grids, targeting cats, were established across the northern plateau and immediately adjacent areas in 2018 and then both the north and south plateaux in 2019. No targeted cat surveys were undertaken in 2020. All cameras deployed within the cat grids were located at a subset of northern quoll grid camera locations. Each grid consisted of 70 cameras, a mix of HC600 Hyperfire Full Covert IR (Reconyx Inc.) and PC800 Hyperfire Professional IR (Reconyx Inc.), at a grid spacing of 500 m × 500 m. Each camera was attached to a tree at a height of 50 cm above the ground. Each camera was pointed towards the same point as the baited (northern quoll) cameras, approximately 2 m away, at an angle of approximately 20° below horizontal. Even though these cameras were not specifically baited to attract cats, deployments often occurred simultaneously with the baited-camera surveys targeting northern quolls or had residual scents from prior baiting surveys. Previous research has not shown an effect of food odours in promoting feral cat visitation (Read et al. 2015; Stokeld et al. 2015), with dried meat being a less receptive bait option for feral cats (Risbey et al. 1997; Moseby et al. 2011). With this consideration, we consider it unlikely that this overlap in bait deployment had a significant impact on our detection frequency of cats. Cameras were deployed for a minimum of 30 days in April–May and October–November of each year. All 70 cameras deployed within each plateau grid were operated simultaneously. Between plateau grids, there were slight differences in camera operation times owing to the time restraints of deployment. They were programmed to take three image bursts per trigger with high sensitivity and 24-h activity. In total, 15 228 trap nights were sampled over 2 years from 140 camera sites across the two plateaux.

The placement of each camera grid was targeted atop the plateaux to prioritise surveying habitat that was inhabited by both species. We aimed to maximise our detections of both species across the plateaux to ensure a thorough analysis of each species’ use of the habitat conditions present. Permission for animal use was authorised by the Charles Darwin University Animal Ethics Committee (A18017) and the Queensland Department of Environment and Science (Scientific Purposes Permit WA0009373).

We conducted ground-based vegetation surveys at each camera-trap site across the two plateaux in April–May 2019. We measured the basal area of all living and dead trees with a diameter greater than 15 cm by using a basal-area sweep with a rod relascope (Wenger 1984; Oakwood 1997). Three sweeps were conducted along a 100-m transect to calculate the tree basal area (m² ha⁻¹) of the stand at each camera site. From this dataset, we calculated the total tree basal area and the basal area of E. tetrodonta, given that E. tetrodonta is the species most frequently used as a den tree by the northern quoll (G. J. Trewella, unpubl. data) for each camera-trap site.

We also derived several variables from satellite products. To measure the productivity and extent of vegetation, we calculated an average normalised-difference vegetation index (NDVI) for each camera-trap site by using 3 years (2018–2020) of Sentinel-2 satellite imagery collected from the months of June and July (European Space Agency 2021), so as to produce an average value for each site. To quantify the the fire regimes of the region, we calculated an average annual fire frequency (fire years⁻¹) and time since last fire (days) at each camera-trap site by using 4 years of high-resolution (20 m) Sentinel-2 satellite imagery (European Space Agency 2021) from all useable images since the inception of Sentinel-2 in 2015. To measure the prominent topographical attributes of the landscape, we calculated elevation and a topographic ruggedness index (TRI; Riley et al. 1999) from a three arc-second digital elevation model (Department of Natural Resources, Mines and Energy 2018a) for each camera-trap site. We also calculated the distance to the nearest water source from surface hydrology line maps (Crossman and Li 2015) and the distance of each camera site to the plateau edge, using a vegetation map (scale: 1:100 000; Department of Natural Resources, Mines and Energy 2018b). In the case of distance to plateau edge, if the camera site was on the plateau, it was given a positive value, and if the camera site was off the plateau, it was given a negative value; hence, the variable ‘distance to plateau edge’ ranged from large negative values (far off the plateau) to large positive values (close to the centre of the plateau).

All statistical analysis was undertaken in R, ver. 3.6.3 (R Core Team 2020). We analysed two response variables, namely, the frequency of habitat use by (1) cats, and (2) northern quolls. The frequency of habitat use was expressed as the proportion of nights (i.e. 24-h periods, from midday to midday) in which the target species was detected.

Two modelling approaches were used to assess the extent to which frequency of habitat use by cats and northern quolls was correlated with environmental variables. Generalised linear modelling (GLM) was used to analyse the frequency of habitat use across all camera sites, by using all environmental variables apart from time-since-fire. Generalised linear mixed-effects modelling (GLMM) was used to analyse frequency of habitat use across all camera sites by using only time-since-fire, with camera site being included as a random effect. Both the GLMs and GLMMs used a binomial error family, given that the response variable was a proportion (i.e. a two-column variable of nights the target species was detected and not detected). All continuous variables were centred and standardised prior to analysis, and we confirmed that there was not excessive collinearity (r ≤ 0.7) among environmental variables (Zuur et al. 2010).

For each species, the 3 years of cat and northern quoll camera-survey detection histories were pooled to create a sole nightly detection/non-detection response variable at each camera-trap site. Because of the differences in camera-survey methods, a nuisance variable termed ‘camera index’, derived from the percentage of cat survey (forward-facing cameras) nights against the total survey nights (forward-facing camera nights and downward-facing camera nights), was included to account for differences in detection probability of each camera-survey method. On the basis of previous unpublished surveys in the region, our understanding is that northern quolls are largely restricted to the bauxite plateaux, whereas cats have been found to move throughout the landscape. As such, we selected slightly different variables to analyse for each species. To model the spatial variation in frequency of cat habitat use, we selected a global model that included the following variables: fire frequency, total basal area, elevation, distance to plateau edge, NDVI, distance to water, topographical ruggedness index and camera index. To model the spatial variation in frequency of northern quoll habitat use, we selected a global model that included the following variables: fire frequency, total basal area, E. tetrodonta basal area, distance to plateau edge, NDVI, distance to water, topographical ruggedness index and camera index. We chose to include elevation as a variable for cat habitat use, to identify whether there was a general preference for elevated habitat (plateau) or low-lying habitat, owing to the observed transient behaviour of cats around the plateaux. We also chose to include the basal area of E. tetrodonta as a variable for northern quoll habitat use because this tree species is a primary den tree for northern quoll in the Weipa region (G. J. Trewella, unpubl. data) and, therefore, a significant environment resource.

The global GLM models of frequency of habitat use by each species were produced using the R package ‘stats’ (R Core Team 2020) and all 128 possible models for each species were evaluated using the R package ‘MuMIn’ (Bartoń 2020). Significant overdispersion was present in the global models; so, QAICc (quasi-likelihood-corrected Akaike’s information criterion) was used to rank models (Burnham and Anderson 1998). To account for the overdispersion, each model was converted to a quasibinomial error structure to estimate model parameters, their standard errors and 95% confidence intervals. Following the information-theoretic approach, all models were averaged to produce conditional parameter estimations (Burnham and Anderson 1998). The sum of Akaike weights for each variable was calculated to indicate relative importance (Symonds and Moussalli 2011).

For each species, the detection history of each camera-trap site per survey over the 3-year study period was recorded as a nightly detection/non-detection response variable. Time-since-fire was calculated from the starting date of each survey at each camera-trap site. If a fire occurred during a survey period, then an additional datapoint was included for that camera-trap site with a time-since-fire value of 1 day. For each species, three models were constructed using the R package ‘glmmTMB’ (Brooks et al. 2017), namely, a null model with a binomial error structure, a single variable (time-since-fire) model with a binomial error structure and a single-variable (time-since-fire) model with a beta-binomial error structure to account for overdispersion. All models included camera site as a random effect. The best fitting model for each species was selected for using the lowest AICc (corrected Akaike’s information criteria).

In total, 274 cat detections were recorded across 27 827 trap nights over the 3-year period. Detections occurred at least once at 135 of 286 camera sites (47%). The mean annual naïve occupancy (i.e. the proportion of sites where cats were recorded) of the northern plateau was 24%. The mean annual naïve occupancy of the southern plateau was 19%.

The frequency of habitat use by cats was clearly correlated with two environmental variables, namely, fire frequency and tree basal area (Fig. 2). These were the only variables present in all well supported models (ΔQAICc ≤ 2; Table 1). Fire frequency had a clear, positive relationship with habitat use by cats; at the least frequently burnt sites (0.25 fires year⁻¹, i.e. one fire every 4 years), cats were very unlikely to be detected (proportion of nights with detection at <0.01 [CI: 0.003–0.007]); at the most frequently burnt sites (1.5 fires year⁻¹, i.e. three fires every 2 years), cats were much more likely to be detected (proportion of nights with detections at 0.04 [CI: 0.02–0.08]; Fig. 3a). Tree basal area had a clear negative relationship with habitat use; at sites with the highest basal area (40 m² ha⁻¹), cats were very unlikely to be detected (proportion of nights with detections at <0.01 [CI: 0.002–0.009]); at sites with the lowest basal area (10 m² ha⁻¹), cats were modestly more likely to be detected (proportion of nights with detections at 0.02 [CI: 0.01–0.02]; Fig. 3b).

Fig. 2.  Conditional regression coefficients of all predictor variables for cat and northern quoll habitat use derived from generalised linear modelling are shown on the left of the solid vertical line, and beta coefficients of time-since-fire for cat and northern quoll habitat use derived from generalised linear mixed effect modelling are shown on the right. Error bars represent the 95% confidence intervals of each estimated coefficient. The dotted horizontal line represents a regression coefficient value of 0.

Table 1.  Parameter inclusion for cat and northern quoll generalised linear models with a ΔQAICc of ≤2.

Fig. 3.  The predicted relationship between the proportion of nights with detections of cats and the most supported predictor variables derived from a model averaged generalised linear model: (a) fire frequency and (b) total basal area. Gray bands represent the 95% confidence intervals of each prediction.

In total, 615 northern quoll detections were recorded across 27 827 trap nights over the 3-year period. Detections occurred at least once at 139 of 286 camera sites surveyed (49%). The average annual naïve occupancy of the northern plateau was 11%. The average annual naïve occupancy of the southern plateau was 38%.

The frequency of habitat use by northern quolls was clearly correlated with five environmental variables, namely, E. tetrodonta basal area, total tree basal area, distance to plateau edge, topographic ruggedness index and NDVI (Fig. 2). These were the variables present in all well supported models (ΔQAICc ≤ 2; Table 1). E. tetrodonta basal area had a clear positive relationship with habitat use; at sites with no E. tetrodonta (0 m² ha⁻¹), northern quolls were unlikely to be detected (proportion of nights with detections at <0.01 [CI: 0.003–0.01]); at sites with the highest E. tetrodonta basal area (25 m² ha⁻¹), northern quolls were an order of magnitude more likely to be detected (proportion of nights with detections at 0.11 [CI: 0.05–0.22]; Fig. 4a). Tree basal area (i.e. including all tree species) had a clear negative relationship with habitat use; at sites with the lowest tree basal area (10 m² ha⁻¹), northern quolls were more likely to be detected (proportion of nights with detections at 0.06 [CI: 0.03–0.10]); at sites with the highest tree basal area (40 m² ha⁻¹), northern quolls were very unlikely to be detected (proportion of nights with detections at 0.002 [CI: 0.001–0.007]; Fig. 4b). Distance to plateau edge had a clear positive relationship with habitat use; at sites located furthest away from the plateau (1600 m), northern quolls were unlikely to be detected (proportion of nights with detections at 0.002 [CI: 0.001–0.01]); at sites close to the centre of the plateau (1400 m from the plateau edge), northern quolls were more likely to be detected (proportion of nights with detections at 0.07 [CI: 0.03–0.13]; Fig. 4c). Ruggedness had a clear positive relationship with habitat use; at flat sites with no ruggedness, northern quolls were unlikely to be detected (proportion of nights with detections at 0.01 [CI: 0.01–0.02]); at the most rugged sites, northern quolls were more likely to be detected (proportion of nights with detections at 0.07 [CI: 0.03–0.14]; Fig. 4d). NDVI had a clear positive relationship with habitat use; at sites with the lowest index value (0.42), northern quolls were unlikely to be detected (proportion of nights with detections at 0.008 [CI: 0.003–0.02]); at sites with the highest index value (0.74), northern quolls were more likely to be detected (proportion of nights with detections at 0.04 [CI: 0.02–0.07]; Fig. 4e).

Fig. 4.  The predicted relationship between the proportion of nights with detections of northern quolls and the most supported predictor variables derived from a model averaged generalised linear model: (a) E. tetrodonta basal area, (b) total basal area, (c) distance to plateau edge, (d) terrain ruggedness index and (e) normalised-difference vegetation index. Gray bands represent the 95% confidence intervals of each prediction. Note that in c, negative numbers denote sites away from the plateau, and positive numbers denote sites atop of the plateau.

Habitat use by cats and northern quolls was best explained by beta-binomial models including time-since-fire. For both species, the beta-binomial model containing time-since-fire had an AICc score >10 less than did the binomial model including time-since-fire and the null model. Cats did not show a clear relationship with time-since-fire, but trended towards a negative relationship (Fig. 2). In comparison, northern quolls showed a clear positive relationship with time-since-fire (Fig. 2).

Remnant populations of the northern quoll on Cape York Peninsula have long been exposed to the main factors thought to be threatening small mammals across northern Australia. Habitat conditions on the bauxite plateaux where these remnant populations occur are likely to have facilitated the persistence of northern quolls despite multiple threats, particularly predation by feral cats, which are locally abundant. Our findings highlighted a nexus among northern quolls, cats and fire, whereby naturally occurring fire refugia (in this case on the less frequently burnt plateau surfaces) benefit northern quolls by promoting structurally complex vegetation and, consequently, limiting habitat use by cats. This is consistent with a growing body of evidence showing that high fire frequency can promote cat activity and abundance, resulting in the decline of native mammals susceptible to cat predation (McGregor et al. 2015, 2016; Davies et al. 2020; Stobo-Wilson et al. 2020b).

Cats appear to be associated with frequently burnt and structurally simple habitat at our study site. Our modelling showed that fire frequency had the strongest influence on habitat use by cats, with cats tending to use frequently burnt areas. Fires are most frequent in the surrounding low-lying open savanna, with the vegetation transition along the rugged edges of the plateau (from open savanna to dense E. tetrodonta savanna) acting as a natural barrier to fire (Bowman 1986). Working in the savannas of the Kimberley and Cape York Peninsula regions, McGregor et al. (2015, 2016) showed that feral cats have improved hunting efficiency in open, grassland habitat and actively seek out recently burnt habitat, often far outside their established home ranges. We found that sites with the highest tree basal area had the lowest frequency of habitat use by cats. Given the findings of McGregor et al. (2015, 2016), it seems likely that the dense plateau vegetation limits the hunting efficiency of cats and, as a result, the suitability of these areas as habitat for cats. In addition to the immediate reduction in vegetation cover caused by fire, repeated fires at short intervals can reduce the structural complexity of tropical savanna communities over the long term (Williams et al. 1999; Russell-Smith et al. 2012; Woolley et al. 2018). Cats are likely to benefit from fire-driven reductions in the density and structural complexity of understorey vegetation (Davies et al. 2020; Stobo-Wilson et al. 2020b) to the detriment of small native mammals (Leahy et al. 2015; McDonald et al. 2016).

In contrast to the use of structurally simple habitat by cats, northern quolls most frequently used structurally complex, long-unburnt habitat. Our modelling found vegetation composition, ruggedness and time-since-fire had the strongest influence on habitat use by northern quolls. Northern quolls most frequently used the central areas of the plateaux; E. tetrodonta basal area, NDVI and ruggedness were also significant positive predictors of northern quoll habitat use. E. tetrodonta is known to be the primary den tree used by the northern quoll across the plateaux (G. J. Trewella, unpubl. data), and the strong association between northern quolls and E. tetrodonta is most likely related to the high rate of hollow formation in this tree species, compared with other savanna trees (Werner and Prior 2007; Woolley et al. 2018). However, habitat use by northern quolls was negatively related to total tree basal area (i.e. including species other than E. tetrodonta), with northern quolls infrequently using areas with high total tree basal area. Without considering the proportion of hollow-bearing trees in the landscape, high tree basal area alone may not provide necessary resources such as food and shelter to support a northern quoll population in habitat lacking other den resources, such as rock crevices. Similar to the findings of Moore et al. (2019), vegetation productivity, inferred from NDVI, was also correlated with habitat use by northern quolls, with the most productive areas being used more frequently by northern quolls. The positive relationship between habitat use by northern quolls and ruggedness is consistent with a continent-wide pattern of remnant northern quoll populations being associated with rugged areas (Moore et al. 2019). Unlike for cats, we did not find a relationship between fire frequency and habitat use by northern quolls, but there was a strong association between northern quolls and long-unburnt habitat. Long-unburnt habitat in tropical savanna is likely to provide a greater cover from predation and have increased availability of denning resources such as hollow logs (Legge et al. 2008; Shaw et al. 2021), while also being avoided by cats (McGregor et al. 2014). These observed differences in habitat usage by northern quolls and cats highlight the importance of less frequently burnt refugia (such as bauxite plateaux on Cape York Peninsula) for the northern quoll across northern Australia.

A resource that appears to be a key determinant of persistence of northern quolls is den availability. Northern quolls are the largest marsupial that displays partial semelparity, with males dying following the annual breeding season (Dickman and Braithwaite 1992; Oakwood et al. 2001). This regular high population turnover means that annual recruitment and juvenile survivorship are vital to population persistence (Cremona et al. 2017; Moro et al. 2019). Ample availability of high-quality den resources (e.g. rock crevices, hollows in large trees and logs), that offer protection from threats during the vital joey-rearing phase, are likely to play a critical role in successful recruitment, and hence population persistence (Oakwood 2000). Where there is a choice of den resources, such as rock crevices or tree hollows, northern quolls have been reported to show a preference for rock crevices (Begg 1981; Oakwood 1997). As such, rocky refugia are seen as more suitable habitat for northern quoll, which can support higher quoll densities than can lowland savanna (Begg 1981; Schmitt et al. 1989; Oakwood 2002). However, savannas that lack rocky features can still provide sufficient den resources if large hollow-bearing trees, particularly hollow-forming species, such as E. tetrodonta (Woolley et al. 2018), are abundant (Thomas et al. 2021). Reducing the frequency of high-intensity fires in tropical savannas is likely to promote the retention of large, hollow-bearing trees, which provide suitable den sites for northern quolls (Williams et al. 1999; Edwards et al. 2018; Woolley et al. 2018). Fire management strategies should therefore be targeted at maintaining structurally complex habitat, particularly through the retention of large hollow-bearing trees.

Despite demonstrating links between habitat conditions and patterns of habitat use by cats and northern quolls, we have not considered how their habitat use is likely to be affected by other species, most notably those whose abundances may have changed significantly in recent decades. Unfortunately, a lack of historical and contemporary surveys of mammals on Cape York Peninsula hinders our ability to identify trends in species distributions and abundances in response to changes in land-management practices and biotic and abiotic processes through time. For example, we have not taken account of the potential role dingoes play, as apex predators, in regulating habitat use by cats and northern quolls (Glen et al. 2007; Hernandez-Santin et al. 2016; Cremona et al. 2017). Dingoes are common across the savannas of western Cape York Peninsula (Cairns et al. 2018); however, persecution by pastoralists, through poisoning and shooting, does occur. Whether this level of dingo persecution is enough to release cats from predation or competitive pressure, as predicted by mesopredator release theory (Ritchie and Johnson 2009; Kennedy et al. 2012), is worthy of consideration. Similarly, cat activity has been found to be highest in savannas where invasive herbivores are most abundant (Davies et al. 2020). Cattle pastoralism is a key industry on Cape York Peninsula; so, the simplification of vegetation through overgrazing is a widespread threat (Kutt and Woinarski 2007). Potential synergies between cats and livestock (both feral and managed) may be an important barrier to mammal conservation across Cape York Peninsula and elsewhere in northern Australia (Stobo-Wilson et al. 2020a).

Research over the past few decades has shown that northern Australian mammals are rapidly declining in response to a variety of threats, particularly predation by feral cats (Woinarski et al. 2011; Frank et al. 2014; Stobo-Wilson et al. 2020a). Given that the eradication of the cat is not currently feasible across unfenced landscapes, the maintenance of complex habitat may be one of the few ways to mitigate cat impacts on native mammals across northern Australia. Fire regimes appear to be a crucial factor determining whether populations of small mammals are able to persist in the extremely fire-prone savannas of northern Australia (Griffiths and Brook 2014; von Takach et al. 2022). Reductions in the frequency of fires, especially those of high intensity, can promote the persistence of large hollow-bearing trees (Williams et al. 1999; Edwards et al. 2018; Woolley et al. 2018) and potentially suppress cat activity (Davies et al. 2020; Stobo-Wilson et al. 2020b). However, achieving long-term reductions in fire frequency and intensity across the vast tropical savanna landscapes of northern Australia will also be a significant management challenge (Murphy and Russell-Smith 2010; Evans and Russell-Smith 2020; Russell-Smith et al. 2020). Prescribed burning in the early dry season (when fire-weather conditions are relatively benign compared to the late dry season) has been shown to be effective in reducing fire intensity (Price et al. 2012; Duncan et al. 2015; Radford et al. 2020, 2021), but typically does not reduce fire frequency. Hence, frequent prescribed burning can still negatively affect small mammal populations and promote cat activity (Andersen et al. 2005; Griffiths et al. 2015). These difficulties in fire management further emphasise the significance of bauxite plateaux as naturally occurring, less frequently burnt refugia, similar to rocky refugia commonly found across northern Australia (Oakwood 2000; Hohnen et al. 2016; Moore et al. 2022). Despite these challenges, retaining structurally complex vegetation and suppressing cat activity through a low-frequency fire regime that promotes the retention of long-unburnt habitat is likely to have a significant benefit for northern quoll and threatened mammals across the tropical savannas of northern Australia.

The data that support this study will be shared upon reasonable request to the corresponding author.

The authors declare no conflicts of interest.

This study was funded by Rio Tinto Weipa (RTA Weipa Pty Ltd) and an Australian Government Research Training Program (RTP) Scholarship.