Spatial Capture-Recapture
We used Spatial Capture-Recapture with the R package oSCR (Sutherland et
al. 2016) in RStudio (RStudio Team 2018) to examine white-tailed deer
habitat selection i.e. how habitat affects white-tailed deer densities
(D) and capture probabilities (p0). We first chose the
top homogeneous model without habitat covariates by AIC and continued
fitting habitat covariates into the parameters of that model. This
simplified the model selection procedure by reducing possible
combinations of covariates (Efford and Fewster 2013, Brazeal et al.
2017). We constructed 42 different homogeneous models where we let D,
p0 and σ (the spatial scale parameter) vary by year,
sampling occasion, and sex. Supplementary Table S3 shows the
combinations of fitted inhomogeneous models where D and
p0 vary by different habitat covariates. All
inhomogeneous models are modifications of the most supported homogeneous
model. We also compared the overall predicted density estimates of the
top homogeneous model with the top inhomogeneous SCR model including
habitat covariates. We used multi-session SCR models with sampling year
as a “session”. The state space was defined by a grid with resolution
120 m, which is about 0.5 x σ based on the estimate of σ from this study
(see Results). A state space buffer of 1000m around the traps was used.
Habitat covariates were defined using the open-source Corine Land Cover
data (2012) provided by Finnish Environment Institute. For water bodies,
we used the vector data of waterways provided by National Land Survey of
Finland (2018). We considered three different habitat covariates for
density. First was a categorical habitat class variable with four
different levels: agricultural areas (fields), coniferous forests, mixed
forests and transitional woodland/shrub. Other density covariates were
distance to artificial areas (e.g. buildings, roads and other
artificially surfaced areas) and distance to water. These covariates
were assigned to the state space by extracting them from the Corine Land
Cover raster data with a function extr.rast() (oSCR package) using the
habitat which was the most frequent when summarizing the Corine Land
Cover raster values (resolution 20m x 20m) around the central
coordinates of the state space pixel on the same resolution as the state
space is defined (here 120m). Artificial areas would have covered only
2% of the state space and those pixels were assigned to the habitat
that was the second most frequent. Artificial areas and water bodies
were included in the analysis by calculating the nearest distance from
the state space pixel central coordinates to artificial area or water.
We included four different trap-level covariates to study how landcover
type affects capture probability (p0). The first
covariate was a categorical landcover class variable with three
different levels: coniferous forest, mixed forests, transitional
woodland/shrub (hereafter: transitional woodland). Other three
covariates for capture probability were distance to agricultural areas,
distance to artificial areas, and distance to water bodies. To define
the landcover class for each plot, the central coordinates of the
sampling plots (“traps”) were buffered by 30m buffer and the
proportion of each landcover type was calculated for the buffered area.
The landcover type with the largest proportion in the buffered area was
assigned as the landcover class for that plot. If two or more landcover
types existed in exactly same proportions, then the class was defined
using field notes on the landcover type of the central coordinate of the
sampling plot. Only 3% of the 92 plots would have been assigned to
deciduous forests. Because of this small proportion, this landcover
class as trap-level covariate was changed to the class that covered the
second largest proportion of the buffer area. To include agricultural
areas, artificial areas and water bodies as covariates on capture
probability, we calculated the nearest distance between the central
coordinate of the plots and these landscape features.