Study design
Forest Fire Management Victoria conducted a prescribed burn within our
study area in mid-May 2019. The fire affected 292 ha, of which 246 ha
was forested and 46 ha was heathland (DEECA 2020). The majority of the
fire (66%) burned at low-medium severity (DELWP, 2020), although
severity was not quantified for heathland areas. We conducted five
repeated mammal surveys using camera traps with infrared flash (Reconyx
HF2X) across 30 sites (Figure 1) over a 12-month period. Surveys were
conducted at six- and two months pre-fire and two weeks, three months,
and six months post-fire, with each survey period being approximately
two months in length (Appendix S1: Table A1). Using a Geographic
Information System, we determined camera trap locations by positioning a
900 m grid over the study area and placing 30 survey points at grid
intersections (ESRI 2014). During deployment, some camera traps were
moved up to 150 m from grid points to account for access issues or to
target nearby game trails or old vehicle/walking tracks, as feral cat
and fox detectability is generally higher on trails compared to off
trails (Geyle et al., 2020). We did not place cameras on public roads or
heavily-used walking trails to reduce the risk of theft.
Of the 30 camera trap sites, 40% (12/30) were burnt during the
prescribed burn (Appendix S1: Table A1: Figure 1), and the mean area
burnt within a 100 m radius (i.e., Fire extent, Table 1) for these sites
was 54% (range 34–95%); highlighting the patchy burning style that is
common of prescribed fires in our study region (e.g., Sitters et
al. 2015; Hradsky et al. 2017a). The burnt and unburnt sites
were not spaced far enough apart to be considered independent for all of
our study species (i.e., they were within the feasible movement range of
some species). Therefore, our study design would be more appropriately
described as a quasi-BACI (before-after, control-impact) design,
acknowledging this potential for spatial dependence between the sites.
At each site, cameras were attached to a tree at a height of
approximately 40 cm facing a lure station two metres away. Each lure
station was comprised of wadding soaked in tuna oil encased in a
polyvinyl chloride (PVC) vent cowl. Lure stations were pegged securely
into the ground, and vegetation in each camera’s line of sight was
cleared to prevent false triggers and to ensure animals were clearly
visible. Cameras were set to record three images per trigger at
medium-high sensitivity, with no delay between trigger events.
Statistical
analyses
Images were processed using CPW Photo Warehouses (Ivan and Newkirk
2016). Animals were identified to species level where possible,
otherwise, they were categorised according to the finest
taxonomic/functional group possible (e.g., ‘unknown small mammal
species’). Each photo sequence was treated as a single point in time and
a detection event was defined as images of the same species on the same
camera that were separated by at least 60 minutes. Species detection
matrices were created using the camtrapR package (Niedballa et
al. 2016) in R version 4.2.2 (R Core Team 2022).
To test the influence of the fire, habitat, anthropogenic, and prey
variables (Table 1) on mammal activity, we fitted generalised linear
mixed models (GLMMs) to each species/group with sufficient data
(Appendix S1: Table A2). There were many zeros (i.e., days in which a
species was not detected) in the detection matrices due to long
intervals between detection events. To account for this, we defined the
response variable as the number of days a species was detected in each
survey period relative to the number of days it was not detected. There
were four species included in our analyses: the red fox, feral cat,
swamp wallaby, and eastern grey kangaroo. To fit models and test our
predictions on smaller mammals (<2,000 g), we pooled
detections from small mammals (<800 g) and medium-sized
mammals (800 – 2,000 g). The species comprising these two groups (see
Appendix S1: Table A2) were recorded too infrequently to fit models to
individual species, and many detections were not identifiable to species
level. We conducted all model fitting and verification using the glmmTMB
(Brooks et al. 2017), MUMIn (Barton 2022), and DHARMa (Hartig
2022) packages in program R version 4.2.2 (R Core Team 2022).
Before testing the covariates for each species/species group, we
constructed models to test the effect of two possible detection
covariates, namely camera placement (on or off trail) and age of lure
(Table 1) on each response variable. Camera placement can influence the
detectability of cats and foxes (Geyle et al., 2020), while the age of a
lure might impact mammal activity either through reduced potency over
time or behavioural alterations (Frey et al. 2017; McHughet al. 2019). While the five survey periods were similar in
length (refer to Appendix S1: Table A1), there was inconsistency in the
timing of lure replacement. Lures for surveys four and five were
replaced part-way through the survey periods, unlike those in surveys
one, two, and three, which were replaced at the beginning. These
detection models incorporated the main effects of both camera placement
and lure age, along with random effects of Survey period and Site,
allowing us to account for repeat sampling over time and any
camera-level variability. We assessed the output of these models and
included camera placement and/or lure age as fixed effects in subsequent
analyses if the 95% confidence intervals did not cross zero.
To test the effect of our remaining variables on mammal activity, we
fitted binomial GLMMs containing three-way interactions between
Treatment (CI), Before-After (BA) and each of the remaining
non-detection covariates (Table 1). These models included a total of 14
variables: two Fire History variables, three variables relating to the
2019 Prescribed Fire, two Vegetation variables, one Topography variable,
two variables representing proximity to Anthropogenic Features, and
potentially one or both of the Detection variables if they influenced
the activity of the species/group (Table 1). We included both the Large
mammal and Small mammal Prey Activity variables (Table 1) for the fox,
and the Small mammal Prey Activity variable for the cat. We used the
‘dredge’ function from the MuMIn package for model selection. This
function only allows a maximum of 31 variables in the global model,
including interaction terms. Due to the complexity arising from fitting
three-way interaction between BA, CI, and most of the aforementioned
variables, we fitted two unique ‘sub-global’ models containing different
sets of variables, each with <31 terms. We then used the
‘merge.model.selection’ to combine the two model selection tables per
species/group and reranked the models by AICc. The selection criteria
for well-supported models were based on a delta Akaike Information
Criterion (ΔAICc) of less than 2 (Burnham and Anderson 2004).
For the Fire extent variable, we fitted a simplified two-way interaction
with BA. This is because all unburnt sites had a Fire extent of 0%,
making it redundant to include the CI variable. We did not fit an
interaction between BA or CI and Vegetation Type, Time Since Fire, or
Fire Frequency, as this resulted in model convergence issues. Moderately
and highly correlated variables (i.e., Pearson’s r ≥0.5) were not
included in the same model, and we excluded NDVI and TPI (Pearson’sr = 0.49) from appearing in the same model because preliminary
analyses showed that well-supported models containing both variables
reported influential interactions that were not supported by the
underlying data. We included the random effects of Site and Survey
Period as per the initial detection models, and we scaled and centred
each of the continuous variables prior to modelling. We limited the
maximum number of variables per model to 10 to avoid issues associated
with overfitting, which meant that only one three-way interaction could
feature in any given model.
Table 1. Descriptions of the predictor variables included in
our generalized linear mixed models of mammal activity in the eastern
Otway Ranges, Victoria. The spatial data from the 2019 prescribed fire
was sourced from DELWP (2020).