Williams-Climate-change-refugia-for-terrestrial-biodiversity_0

more prone to shift their ranges to track favourable climatic conditions rather than to
remain in place and evolve new forms (Parmesan 2006).
The climate variables used in this study were: (i) annual mean temperature; (ii)
temperature seasonality; (iii) maximum temperature of the warmest period; (iv) annual
precipitation; (v) precipitation of the driest period; vi) precipitation of the wettest period;
and (vii) precipitation seasonality. Presence-only modelling methods can be subject to
sampling bias (Yackulic et al. 2013); we account for this potential bias by using a
target-group background, which consisted of the locations of the occurrence records for
the species within that class, as recommended by Phillips et al. (2009). Using the
target group as our background points, it is assumed that any sampling bias in our
occurrence records for a single species can also be observed in our background
points; in effect cancelling out the impact of any spatial sampling bias in the modelling
exercise (Phillips and Dudik 2008, Elith and Leathwick 2009, Phillips et al. 2009).
Species distribution models were projected onto future scenarios consisting of the
worst-case scenario RCP8.5 and 18 GCMs for 2085. This RCP was used for the same
reason it was used for examination of future climate. Namely, the current trajectory of
radiative forcing and emissions are most closely aligned with this RCP, and by using
the worst-case scenario we have covered the spectrum of outcomes, as the difference
between the RCPs is the severity of change, rather than the direction of change.
Across the 18 GCMs, the 10 th , 50 th and 90 th percentiles were calculated to give the
median projection and a measure of across-model variance (a measure of uncertainty
in our projections). Due to the large number of species used in this analysis, we
focussed on the shifting climate space of potential distributions by assuming that all
species had unlimited dispersal. This assumption is clearly unrealistic, but
interpretation of our results (section 3.4) is not dependent on the assumption being
true.
The default Maxent distribution output is a continuous prediction of environmental
suitability for the species. The output species distribution models for current climate
were vetted by comparing the predicted distribution model to the published distributions
(generally equivalent to the Extent of Occurrence based on a minimum convex
polygon) of the species: again, taken from relevant field guides (Menkhorst and Knight
2001, Churchill 2008, Tyler and Knight 2009, Wilson and Swan 2010, Vanderduys
2012), online databases (http://www.arod.com.au/arod/) and from expert opinion. A
binary distribution output was created by applying an appropriate threshold obtained
from the Maxent results output file. Two Maxent-generated thresholds were trialled:
‘equate entropy of threshold and original distributions logistic threshold’ and ‘Maximum
training sensitivity plus specificity logistic threshold’ for each species. The one best
representing the known distribution of the species was used (APPENDIX 3. Species
distribution modelling data and model results). The same threshold was used for the
species current and future distribution projections. Applying this threshold generates a
map of where we realistically expect the species to be under a given climate.
Model performance was evaluated by the area under the receiver operating
characteristic curve (AUC). AUC measures each model’s consistency and predictive
accuracy (Ling et al. 2003). An AUC score of 1 is a perfect model fit of the data; 0.5 is
no better than random (Elith et al. 2006, Phillips et al. 2006). AUC values ≥ 0.7 indicate
‘useful’ models, whereas values ≥ 0.9 indicate models with ‘high’ performance (Swets
1988). Models for each species were screened for low AUC (