Movement and habitat selection of a large carnivore in response to human infrastructure differs by life stage

Background The movement extent of mammals is influenced by human-modified areas, which can affect population demographics. Understanding how human infrastructure influences movement at different life stages is important for wildlife management. This is true especially for large carnivores, due to their substantial space requirements and potential for conflict with humans. Methods We investigated human impact on movement and habitat selection by GPS-collared male brown bears (Ursus arctos) in two life stages (residents and dispersers) in central Sweden. We identified dispersers visually based on their GPS locations and used hidden Markov models to delineate dispersal events. We used integrated step selection analysis (iSSA) to infer movement and habitat selection at a local scale (availability defined by hourly relocations), and resource selection functions (RSFs) to infer habitat selection at a landscape scale (availability defined by the study area extent). Results Movement of residents on a local scale was facilitated by small forestry roads as they moved faster and selected areas closer to forestry roads, and they avoided areas closer to larger public roads and buildings on both scales. Dispersers were more ambivalent in their response to human infrastructure. Dispersers increased their speed closer to small forestry roads and larger public roads, did not exhibit selection for or against any road class, and avoided areas closer to buildings only at local scale. Dispersers did not select for any features on the landscape, which is likely explained by the novelty of the landscape or their naivety towards it. Conclusion Our results show that movement in male brown bears is life stage-dependent and indicate that connectivity maps derived from movement data of dispersing animals may provide more numerous and more realistic pathways than those derived from resident animal data alone. This suggests that data from dispersing animals provide more realistic models for reconnecting populations and maintaining connectivity than if data were derived from resident animals alone. Supplementary Information The online version contains supplementary material available at 10.1186/s40462-022-00349-y.


Fitting HMMs
In the following section, we fit three models with two behavioral states, representing short and long movement patterns. Each set of initial parameters are different, which increases the chance of finding the global maximum (citation).
Fitting a hidden Markov model to movement data requires four starting parameters specified for n behavioral states. The first parameter is mean step length distance, mu. The second is the standard deviation for each mean, sigma. The third parameter is the mean turning angle, angleMean. The final parameter is kappa, concentration of the turning angle.
The first set of initial parameters were chosen by looking at summaries of the step lengths and turning angles, histograms and density plots. The second set of initial parameters were "wider" (smaller/larger) values than the first set. The initial parameters for the third model were the estimated parameters from the previous model.
The model specification is as follows: fitHMM(move_object, number of behavior states, initial parameters, theoretical distribution for step lengths, theoretical distribution for turning angles) We selected the gamma distribution for the step lengths and the vonMises distribution for the turning angles.

Histogram of W0612_2a_res$angleRes
We fit the remaining two models with different initial parameters.
Fitting of the 3-state HMM is the same as for a 2-state except that we have three values for each parameter that represent the behavioral states. Initial parameter values for the 3-state HMMs were chosen as follows: 1. Educated guess based on step length and turning angle summaries. 2. Creating three equal bins for the distributions. 3. From a combination histogram-density plot in ggplot2. 4. The estimated parameters from the previous model.

Model selection and assigning behavioral states to steps
We use Akaike's Information Criterion (AIC) to determine which model has the best fit to the data.

Diagnostic plotting of behavioral states
We constructed a dataframe that summarizes the states per day before plotting.
allBx <-summaryBy(state~date + state + ID, FUN=length, data = W0612Reg) prop <function(x) x/sum(x) allBx <-ddply(allBx, "date", transform, share = prop(state.length)) allBx$state <-as.factor(allBx$state) We plot the data in ggplot to look at how the proportion of time in each state changes over time. We look only at the second and third behavioral states, as resting is not important in determining whether or not dispersal is occurring.