#Example 1: Examining European Bird Species’ Range Losses and
Climate Change
In the first example, we explored how climate change metrics relate with
the loss of bird species’ ranges across Europe. For this purpose, we
used distribution data for three European bird species (Milvus
milvus , Cettia cetti , and Merops apiaster ) that were
digitized from published resources. The data included species
presence-absence at a 50x50 km spatial resolution for two time periods
obtained from European Breeding Bird Atlas available for two time
periods (EBBA1 in the 1980s (Hagemeijer et al., 1997) and EBBA2 in
(Keller et al., 2020). Pixel level population losses were calculated by
comparing species records between the two time periods.
Monthly time series of climate variables, such as maximum and minimum
temperature and mean annual precipitation, were obtained from the
updated version of the Climate Research Unit’s database. These climate
variables covered the period between 1957 and 2017, including twenty
years before the first atlas (EBBA1).
Monthly time series of climate variables including maximum and minimum
temperature and mean annual precipitation were obtained from the updated
version of the Climate Research Unit’s database
(http://www.cru.uea.ac.uk). These climate variables covered the
period between 1957 and 2017, including twenty years before the first
atlas (EBBA1) on the assumption that species’ ranges respond to the
long-term climate conditions.
The climetrics R package was used to characterize various
dimensions of climate change between 1980 and 2017. We then tested
whether and to what extent the climate change metrics explained species
range losses.
Four machine learning algorithms were employed to characterize the
relationship between the species loss (as the response variable) and
various metrics of climate change (as the predictor variables), and then
measure the relative importance of climate change metrics to ‘explain’
species range loss. The algorithms include Generalized Linear Models
(GLM), Generalized Additive Models (GAM), Random Forests (RF), and
boosted regression tree (BRT). We used the sdm R package (Naimi
& Araújo, 2016) for modelling species distributions. In order to avoid
biases to the parameter estimation and measure performance of the
models, we used a bootstrapping resampling approach (e.g., Hastie et
al., 2009) implemented in the sdm R package with 50 replications
for each species and modelling method.
The fitted models were used to predict climate suitability loss over the
study area (e.g., Naimi et al., 2022). For each species, considering the
four modelling methods and 50 replications, a total of 200 projections
of climate suitability for were obtained for each species considered. We
then used an ensemble approach (Araújo & New, 2007) to combine these
predictions into a single layer using a weighted averaging function
implemented in the sdm R package. We also calculated the relative
importance of each climate change metric in explaining climate
suitability loss for each bird. The metrics included extreme climate
events, standardized local anomalies, climate change velocity, trend in
precipitation, trends in temperature, average temperature, and
precipitation in spring on the assumption that species respond to spring
weather conditions (e.g., Ahola et al., 2004; Kluen et al., 2017).
We used AUC (area under the curve of a receiver operating characteristic
of extinction events) for evaluating model performance in predicting
extinction events of the three birds considered (Fielding and Bell,
1987) (Fig.2).