Melville Island is situated approximately 25 km off the central coast of northern Australia, and is Australia’s second-largest island (5,786 km²) (Figure 1). Melville Island is the larger of the two main Tiwi Islands, a critical refuge for numerous mammal species that are in severe decline on the northern Australian mainland. However, recent evidence shows that mammals have declined on Melville Island over the last two decades (Davies et al., 2018). Although the declines have not progressed to the same extent as the mainland, the same processes are likely impacting the island populations (Stobo-Wilson et al., 2020a).

Melville Island experiences a tropical monsoonal climate, with over 90% of rainfall occurring in the wet season (Australian Bureau of Meteorology, 2020a). There is a substantial gradient in mean annual rainfall across the island, from 1,400 mm in the east to 2,000 mm in the north-west. Tropical savanna, dominated by tall eucalypts (predominantly Eucalyptus miniata, Eucalyptus tetrodonta, and Corymbia nesophila), covers 75% of the island (Woinarski et al., 2003). The understory is predominantly grassy, and shrub density is highly variable. Due to the high density of large trees, Melville Island generally supports a higher tree hollow densities than the adjacent mainland (Woolley et al., 2018; Penton et al., 2020a). Although hollow densities increase with mean annual rainfall across Melville Island, there is still substantial localized variability, driven by disturbance, such as tropical cyclones, fire, termite activity (Woolley et al., 2018). These forms of disturbance are particularly important drivers of hollow availability on Melville Island because the region is frequently exposed to severe tropical cyclones (Cook, 2001; Cook and Goyens, 2008), experiences very high fire frequencies (Richards et al., 2012), and has particularly high abundance and species richness of termites (Abensperg-Traun and Steven, 1997).

We focused on three hollow-dependent, semi-arboreal mammals that once occurred across vast areas of northern Australia at the time of European arrival but are now in various stages of decline (Woinarski et al., 2011, 2014). All three are endemic to the tropical savannas. The northern brushtail possum (Trichosurus vulpecula arnhemensis) has declined severely (Stobo-Wilson et al., 2019), and remains abundant only in the highest-rainfall parts of its former range (von Takach et al., 2020), including the Tiwi Islands (Davies et al., 2018, 2021). The black-footed tree-rat (Mesembriomys gouldii) is one of Australia’s largest rodents and has declined severely in both the Kimberley and mainland Northern Territory (Firth et al., 2010; von Takach et al., 2020), while the Melville Island subspecies (M. g. melvillensis) has also shown considerable decline (Davies et al., 2018). Finally, the brush-tailed rabbit-rat (Conilurus penicillatus) has experienced a very severe decline (Firth et al., 2010), and the Tiwi Islands subspecies (C. p. melibius) has displayed the same rapid pattern of decline on Melville Island (Davies et al., 2017). All three species (or subspecies in the case of the northern brushtail possum) are listed as Vulnerable under Australia’s Environment Protection and Biodiversity Conservation Act (1999). Both rodents are listed as Vulnerable on the IUCN Red List (Burbidge and Woinarski, 2016; Woinarski and Burbidge, 2016).

Penton et al. (2020b) recently identified that the brush-tailed rabbit-rat, black-footed tree-rat, and northern brushtail possum predominantly den in large hollows (≥10 cm entrance diameter), which are generally found in large eucalypt trees [≥30 cm diameter at breast height (DBH), 1.3 m]. Large hollows are relatively scarce in northern Australia’s tropical savannas (estimated as only 10% of all hollows in the landscape), and trees with large hollows are particularly vulnerable to disturbance (Murphy and Legge, 2007; Woolley et al., 2018).

In 2015, camera trap surveys were conducted at 88 sites across the lowland savannas of Melville Island (Davies et al., 2017), which had been selected from previous surveys in 2000–2002 (see Firth et al., 2006a). Davies et al. (2017) selected the subset of sites located within savanna woodland, stratified by fire frequency and rainfall, separated by a minimum distance of 1 km. Each site was surveyed with five horizontally facing motion-sensor cameras deployed in a square formation around a central camera, with 50 m between the central and outermost cameras. A bait station consisting of peanut butter, oats, and honey was placed 1.5 m from two of the five cameras, whereas the remaining three cameras at each site had a bait station placed 2.5 m from the camera (Gillespie et al., 2015). Vegetation within each camera’s field of view was cleared to create a fire break and reduce the chance of false triggers, extending battery life. Of the five cameras deployed at each site, two were HC550 Hyperfire white-flash cameras, whereas the remaining three cameras PC800 Hyperfire Professional infra-red flash cameras (Reconyx, Inc., Holmen, WI, United States). Cameras could be triggered at any time of day and were set to a high sensitivity, with cameras programmed to take three images per trigger, with a 1-s interval between images. The cameras were deployed for at least 35 consecutive days (range 35–53 days).

This dataset has been analyzed previously by Davies et al. (2017, 2018), to investigate environmental correlates of the site-occupancy of mammals on Melville Island. However, due to very high site-occupancy rates for both the black-footed tree-rat and northern brushtail possum (Davies et al., 2018), occurring at most sites across the island, and gregarious social structure of the brush-tailed rabbit-rat (i.e., strong spatial clustering of populations; Firth et al., 2006b; Penton et al., 2020b), we considered that abundance was a more appropriate response variable (cf., site occupancy) to test our hypothesis. Hence, we used Royle–Nichols abundance-induced heterogeneity models, which assume that heterogeneity in detection frequency is indicative of variability in a species’ abundance (Royle and Nichols, 2003).

Davies et al. (2017, 2018) investigated a range of environmental correlates of site-occupancy of the brush-tailed rabbit-rat, black-footed tree-rat, and northern brushtail possum. We used similar environmental variables, but to evaluate the influence of the availability of tree hollows and hollow-bearing trees on the abundance of hollow-dependent mammals, we revisited the mammal survey sites in 2017, 2 years after the original survey, and measured two additional variables: (1) hollow density and (2) density of large eucalypts (which are more likely to contain numerous hollows; Woolley et al., 2018; Table 1). Six of the sites surveyed in 2015 could not be resurveyed in 2017 and were excluded from the analysis, leaving 82 sites (Figure 1).

At each site, a random central reference point was selected following the plotless “point centered quarter” method (Krebs, 1999). To increase number of points sampled up to 12 trees (alive or dead) with a DBH of ≥15 cm closest to the central point, equally represented in each quadrant were selected for hollow surveys, which produced a total sample of 963 trees across 82 sites. The hollows in the selected trees were counted from the ground using binoculars. A hollow was defined as having an entrance diameter of ≥5 cm. All ground-based surveys were performed by one individual (CP) and were conducted in good light and weather conditions for optimal visibility (e.g., no rain or mist). A minimum of 2 min was spent surveying each tree for hollows from all available vantage points. We recorded the total number of hollows in two entrance diameter classes (5–10 cm and ≥10 cm) for each tree. Due to the high frequency of termitaria blocking hollows <10 cm entrance diameter, we calibrated 5–10 cm hollows before calculating total hollow counts (≥5 cm and ≥10 cm entrance diameter) using a subset of trees that were double sampled with climbing surveys (Penton et al., 2020a).

We also recorded calculated large eucalypts (dead or alive) ≥30 cm DBH at each site from the original tree dataset, which produced a sub-sample of 465 trees across the 82 sites. We calculated hollow and large eucalypt densities using the plotless “point-centered quarter” method where the area around the random point is divided into four quadrants, and the distance to the nearest tree is measured in each quadrant:

Where n is the number of samples and r_ij is the distance (m) from random point i to the closest individual in quarter j (Krebs, 1999; Jamali et al., 2020).

Penton et al. (2020b) found that den tree selection was significantly correlated with the presence of a large hollow (i.e., ≥10 cm entrance diameter) in that tree. They found that the brush-tailed rabbit-rat used tree hollows with an entrance diameter ≥5 cm. Hence, for this species, we calculated hollow density using hollows with an entrance diameter ≥5 cm (Penton et al., 2020b). For the larger-bodied species (black-footed tree-rat and northern brushtail possum), which tend to use larger hollows, we calculated hollow density using only hollows with an entrance diameter ≥10 cm.

All analyses were undertaken in the statistical program R (R Core Team, 2020). To examine the relationship between the abundance of three hollow-dependent species and our ten environmental variables (Table 1), we used Royle–Nichols abundance-induced heterogeneity models (Royle and Nichols, 2003) in the R package “unmarked” (Fiske and Chandler, 2011). This model assumes that heterogeneity in detection frequency is indicative of variability in a species’ abundance by assuming there is a constant probability of detecting each individual, and detections are independent. We created nightly detection histories for each of the three mammal species at each site by dividing the camera surveys into 24-h sampling occasions. Mammal detections were combined for all cameras at each site.

We developed a set of 1,024 candidate models to explain variation in the abundance of each mammal species. This represented all combinations of the ten environmental variables, with no interactions. Models were ranked and evaluated using Akaike’s Information Criterion (AIC). All continuous environmental variables were log-transformed, then centered and standardized by deducting the mean and then dividing by twice the standard deviation (Gelman, 2008). We confirmed that there was not excessive collinearity among environmental variables by examining the variance inflation factor (VIF). Following Zuur et al. (2010), we used a conservative VIF threshold of 3; if a variable had VIF > 3 we removed it. We further screened collinearity of variables with known associations (e.g., feral cat activity and fire frequency) with Pearson’s correlation coefficient, r, to confirm all variables had a weak or negligible correlation (i.e., r < 0.45). We checked for spatial autocorrelation in the residuals of the global model by visually examining variograms of the model residuals and conducted a Moran’s I test on the residuals of a linear regression global model with a response variable of total nights the species was recorded at a site in R package “ape” (Paradis and Schliep, 2019); brush-tailed rabbit-rat Moran’s I = 0.08, p < 0.001; black-footed tree-rat Moran’s I = 0.05, p = 0.05; northern brushtail possum Moran’s I = 0.08, p = 0.002 indicating very weak spatial autocorrelation (Moran, 1950).

The abundance of arboreal mammals has been shown to have a “humped response” to hollow availability (Wormington et al., 2002). Thus, we initially considered a quadratic function (i.e., humped relationship) between the abundance of the three mammal species and each of hollow density and large eucalypt density. However, there was no evidence that quadratic relationships were more suitable than linear relationships. We also examined interactions between shrub density and large eucalypt density, the inclusion of the interaction terms did not substantially improve the best models (i.e., AIC reduced by ≥2).

We created a global model, including all ten environmental variables, and plotted the standardized coefficients to evaluate the relative effect sizes of each variable (Figure 2). Variables that occurred in all of the well-supported models (ΔAIC ≤ 2; Table 2) were considered important environmental variables.

Due to the higher densities of hollow-bearing trees on Melville Island compared to the mainland (Woolley et al., 2018), we are unable to speculate as to whether the depletion of tree hollows has contributed to the broad-scale decline of arboreal mammals in northern Australia. Frequent late dry-season fires have been shown to impact the density of suitable hollows for the threatened Gouldian finch (Erythrura gouldiae) in northern Australia (Radford et al., 2021). In habitats with low hollow densities, the frequent utilization of alternative dens (e.g., hollow logs, Pandanus spiralis) may expose individuals to greater predation risk and impact breeding success, particularly after fire (Leahy et al., 2016; Penton et al., 2020b). Future research should evaluate how mammal abundance varies with hollow availability across larger spatiotemporal scales to address whether frequent high-intensity fires have contributed to the dramatic decline of hollow-dependent arboreal species in northern Australia. Manipulative experiments using nest boxes could also provide insights into whether an increase in denning resources leads to an increase in native mammal abundance.

It is essential that the conservation management of native mammals in tropical savannas focuses on maintaining a complex and diverse understory. Studies on the northern Australian mainland have shown shrub density is strongly influenced by fire regimes, with frequent, high-intensity fires tending to reduce shrub density (Edwards et al., 2003, 2018; Russell-Smith et al., 2003). The impacts of which are also likely exacerbated by high feral herbivore densities (Legge et al., 2019). There needs to be a greater exploration of how frequent fires, even those of low intensity, influence vegetation structure and diversity, especially in relation to the shrub and grass layers, which provide critical shelter from predators and also provide food resources (e.g., fruits, flowers, and seeds; Bowman, 1988; Fensham, 1990; Paramjyothi et al., 2020). In the relatively intact savannas of northern Australia, such as Melville Island, management should aim to conserve hollow-bearing trees while retaining dense and complex under-and mid-story vegetation, which appears to be essential for the conservation of hollow-dependent arboreal mammal species.