2.2 Predictor variables
We included environmental predictors that encompassed land cover,
climate, and topography attributes measured across the entire state of
Connecticut. We used building footprints as a proxy for development
(Microsoft 2019). We used the 1.0 m resolution Coastal Change Analysis
Program land cover map to extract mixed forests (National Oceanic and
Atmospheric Administration 2023). We used a Young Forest and Shrubland
Vegetation Map to identify different types of young forest and shrubland
vegetation within Connecticut (Rittenhouse et al. 2022). The map used
ecological processes (succession, disturbance, regeneration, hydrology),
vegetation height, percent vegetation cover by height category, previous
land cover type, and time since disturbance to classify different types
of young forest and shrubland vegetation (Rittenhouse et al. 2022). We
also used an understory vegetation map to identify locations of
greenbrier (Smilax rotundifolia ), Japanese barberry
(Berberis thunbergii ), mountain laurel (Kalmia latifolia ),
and other mixed invasive understory species within deciduous forests of
Connecticut (Yang et al. 2023). Using the National Wetlands Inventory
(US Fish and Wildlife Service 2020), we extracted wetlands and a
specific list of forested/shrub wetland types found in sites known to be
occupied by New England cottontail (Rittenhouse et al. 2022). We
calculated elevation, northern and eastern aspects, and slope from 30 m
elevation tiles (National Aeronautics and Space Administration 2000). We
used 1 km resolution annual total precipitation (mm) data (Thornton et
al. 2020) and averaged annual total precipitation across multiple years
within the study period (2016-2019). Fine-scale annual precipitation
data was only available until 2019, so we could not measure total annual
precipitation for more recent years of the study. We removed
hydrological features from all predictors to avoid identification of
lakes, ponds, rivers, and streams as suitable habitat (Elith et al.
2011). To improve the performance of subsequent analyses, we converted
all land cover predictors to continuous predictors by creating distance
surfaces with an offset of 1 to replace all zero distance values, since
zero-inflated data can bias Principal Component Analysis (PCA) results
(Hellton et al. 2021).
We also used a composite layer of the young forest and shrubland
vegetation map and the understory map to run a morphological spatial
pattern analysis (MSPA) using the Guidos Toolbox (Soille and Vogt 2009,
2022, Vogt and Riitters 2017, Vogt et al. 2022). We input the composite
raster into Guidos Toolbox and set the presence of young forests,
shrublands, and understory as the foreground and absence of those as the
background. We used the SPA6 option to classify the foreground into six
categories: core, edge, perforation, fragment, margin, and core opening.
In the foreground, core indicated interior area, edge was the external
object perimeter of the core, perforation was the internal object
perimeter of the core, fragment was disjointed areas too small to
contain core, margin was non-contiguous area that did not fit into the
other categories, and core opening was the area within perforations
(Vogt and Riitters 2017, Vogt et al. 2022). Core opening was the only
background class created not from the presence of the input layer, but
instead from the arrangement of the foreground around it (Vogt and
Riitters 2017, Vogt et al. 2022; Figure 1). The output of the MSPA was
included as a categorical predictor in the habitat suitability models.
Once all predictors of interest were retrieved, we used pair-wise
Pearson correlations to remove highly correlated predictors (Kuemmerle
et al. 2011), r > 0.50. When a predictor was correlated
with multiple predictors, we retained the predictor(s) with the fewest
correlations and the greatest biological relevance. We also used a
Maxent model run with all predictors to ensure we did not remove
predictors that were influential to habitat suitability.