Modelling local habitat suitability
A total of 1083 presence records (all direct observations during visual
surveys), across 10 study locations representing the range of habitats
used by P. muralis (urban, suburban, rural), were used in
producing relative habitat suitability maps and predictive models of
local range expansion. These study sites encompassed heterogeneous land
cover that helped in identifying variables affecting local habitat
suitability and features acting as important corridors for range
expansion. Data for six environmental variables at 2m resolution were
used for the MaxEnt input and are summarised in Table 1. All variables
were calculated and prepared in ArcGIS® (Esri 2017).
We used the Phase One Habitat Survey Toolkit (Centre for Ecology
Environment and Conservation, 2018) to create fine scale habitat type
(categorical) data layers.
Modelling local range expansion (IBM )
Habitat suitability maps from our local scale MaxEnt models were
prepared as habitat quality landscape layers by linear transformation of
the MaxEnt logistic values (estimates between 0 and 1 of probability of
presence) above the maximum test sensitivity plus specificity logistic
threshold. This is the threshold at which the MaxEnt models maximize
their discrimination of presences from background data (Jimenez-Valverde
& Lobo, 2007; Glover-Kapfer, 2015). The resulting habitat quality
landscape (scaled 0-100, and where cell values scale with cell carrying
capacity in RangeShifter), provided the patch input for RangeShifter
v1.1 (Bocedi, Palmer, et al., 2014), in addition to a cost layer to
movement created by reclassifying (inverting) the habitat quality
landscape layer. All inputs were resampled using bilinear interpolation
to 15m x 15m cell size to reduce demands on computational memory whilst
retaining biological relevance to wall lizard movement capabilities. A
single cell in each landscape was identified as the initial species
distribution (i.e., point of introduction for each population
respectively) based on knowledge of the precise location of introduction
when known, or by using the centre point of the current extent of
sighting records for the population.
Parameterisation
Parameters of wall lizard demographics and behavioural attributes were
based on empirical data in the published literature. Where published
empirical data were not available, reasonable judgements and/or
simplifying assumptions were made. The final parameter values used were
biologically realistic and justifiably reflect the functional biology ofP. muralis (Table S2 in Appendix 1). Parameterisation was further
refined through an iterative process, where simulations were repeated
across all study sites with fine parameter adjustments within
biologically meaningful limits until a single set of parameters was
found that modelled as closely as possible the currently observed
spatial extent of each study population (Fraser et al., 2015).
Initialisation
Simulations were initialised using known founder size where documented
(Michaelides et al., 2015; Langham, 2019). Where founder size was
unknown, we used a minimal founder size that resulted in reasonable
simulation outputs as per the iterative process mentioned above. We
assumed adult age class for all founders. Local extinction probability
was set at a constant of 0.003 across populations. Simulations (50
replicates) of population range expansion for the 10 study populations
were then run for the period of time since introduction (which varies
among sites) up to the year 2040.
Analysis
We investigated how landscape characteristics might influence population
size, rate of population growth and range expansion, by first obtaining
standard population growth metrics: carrying capacity (K ), and
intrinsic rate of increase (r ), from linear growth curves applied
to mean yearly population size data taken across all simulation
iterations in R Studio (R Core Team, 2017) using the package
Growthcurver (Sprouffske, 2018). We then created binary habitat
suitability layers from our MaxEnt outputs for a radius of 200m around
introduction points which served as inputs for the programme FRAGSTATS
v4 (McGarigal, Cushman, & Ene, 2002). We ran linear regression models
with two FRAGSTAT metrics describing heterogeneity of suitable habitat
patches within the landscape (Normalised Landscape Shape Index – a
measure of patch aggregation; and Connectance – a measure of functional
joinings of patches) and average habitat quality as explanatory
variables, and the growth rate parameters (k, r ) and
annual dispersal distance as response variables. We set the threshold
distance within which patches are deemed ”connected” to an arbitrary
100m.