Study area
The study area is situated in and around the two wind farms at Storliden mountain (eight wind turbines being constructed in 2010–2011) and at Jokkmokksliden mountain (ten wind turbines being constructed in 2010–2011), located in the calving and post-calving ranges of the Malå forest reindeer herding community. The range used by the reindeer varies between years, it amounts to 870 km2 (excluding lakes), and is approximately 30 % of the entire range used during the snow-free period in the community (Fig. 1). The study area was used by approximately 1200 female reindeer and their calves (the total number of reindeer in the community is 6200 in the winter herd; pers. com. J Rannerud, head of Malå reindeer herding community). It is characterised by undulating forest interspersed with mires, lakes and hills or smaller mountains, with forested land comprising old growth forest, clear cuts and plantations. The calving area is generally considered by the herders to consist of three separate grazing areas for the reindeer as a result of human activity and roads, with specific migration routes or connecting corridors (pers. com. Rannerud; Sametinget 2015). An eastern and a western portion are separated by a public north–south spine road running from the town of Malå to the north, connecting several smaller settlements. The relatively steep mountain slopes of Jokkmokksliden and the hills east of Jokkmokksliden create naturally narrow movement corridors between reindeer foraging areas to the east and the west of the road and the hills. The eastern portion is then, in turn, separated into one northern and one southern part by a main road running from east to west without settlements, and also by a separate 30 kV power line further north stretching from the east to the west through the whole study area, but the latter is not associated with any human movement (Fig. 1). The forest at Jokkmokksliden is intensively managed and contains a high proportion of Pinus contorta plantations (20 %) and dense young forest (30 %), giving little ground vegetation cover. Forest with trees <5 m is classified as young forest, and open areas with trees <2 m are classified as clear cuts (Reese et al. 2003). At Storliden there are areas of old growth forest, plantations of P. contorta (14 %), young forest (20 %), and clear cuts (14 %). According to the herders, the reindeer prefer Storliden just after calving (pers. com. J. Rannerud). Apart from the forest harvest on behalf of the wind farm, there were no harvesting actions in the study area during our study period.
GPS data collection
A total of 80 female reindeer cows have been fitted with GPS collars since 2007 (Followit Lindesberg AB, reindeer collar). As we studied the use of a calving range adult females were chosen as study animals. Furthermore, females also make up the majority (ca. 70 %) of the herd, and they normally move as part of a herd except for about 1 week after individual calving, and are thus likely to provide the most representative picture of how the reindeer herds use their ranges (Skarin and Åhman 2014). We used position data collected every 2 h from female reindeer that occupied the study area in and around the construction site during the calving and post-calving seasons in 2008–2011 (Table 1). The start of the calving season was defined as occurring after the arrival of the reindeer in the study area following spring migration from the winter grounds; the season ends a period of free ranging with the reindeer herders starting to gather the reindeer for calf marking. The exact dates of the start and the end of the season differed between the four study years. Fix success of the GPS collars was sometimes low, ~1 % of the positions were removed having high DOP-value (>10) and 2D positioning (Frair et al. 2010). The data was downloaded from the collars via the global system for mobile communications (GSM) cell phone network, at some occasions the reindeer used areas with low GSM-coverage not making data download possible. Thus, we excluded reindeer with more than 15 % of their positions missing due to holes in the data. In total we used 23,205 locations from 5 to 15 individual reindeer recorded each year in our analysis (Table 1).
Table 1 Study phases in the Malå study area in northern Sweden, over the four study years with number of days and number of reindeer fitted with devices providing a GPS position every 2 h
Habitat variables
Variables assembled to explain reindeer habitat selection in the RSF-models were vegetation type, forest height, elevation, ruggedness, slope, and minimum distance to road, power lines and new infrastructure within the wind farm area. All variables were screened for collinearity using the Variance inflation factors (VIF; Zuur et al. 2009), and we used VIF ≥3.0 as a threshold for removing a variable. Furthermore, pairwise correlation was also checked for using Pearson’s correlation coefficient between the variables, and we used r ≥ 0.5 as the threshold for removing one of the variables. All environmental parameters (Table 2) were first extracted using Arc GIS 9.3™ software (ESRI Inc., © 1999–2009). All the digitized geographical data, except the forest height structure (Swedish kNN-layer; ftp://salix.slu.se/download/skogskarta/), were provided by Lantmäteriet (http://www.lantmateriet.se), which supplies national geographic and land information data. The kNN height data were derived from a combination of satellite imagery, field data and algorithms that validate and determine k-Nearest Neighbour (kNN) distances (Reese et al. 2003). There are 43 vegetation classes from the Swedish Landcover Map present in the area (SMD, Lantmäteriet 2004). We complemented the SMD layer, which originates from satellite images from the year 2000, with satellite data for each study year to include changes from old forest to clear-cuts and clear-cuts to young forest. The 43 classes were then combined into eight classes (Table 2). The class variables were resampled from the 25 m grid to a 50 m grid, where the most common class from the 25 m grid determined the new class of the 50 m grid. The digital elevation model layer provided had a resolution of 50 m and a vertical accuracy of ±2 m. The ruggedness index was calculated as described by Sappington et al. (2007) with a 5 × 5 neighbourhood using the SAGA GIS (http://www.saga-gis.org) module.
Table 2 Proportion, densities or ranges (median value in parentheses) of environmental parameters (50 m resolution) within the Malå study area
Wind farm construction site and existing infrastructure
The years 2008 and 2009 were defined as pre-development years (hereafter referred to as the “pre-construction” phase), during which the area had existing infrastructure such as a road network, power lines and an underground mine on the north side of Storliden. In total, there were 265 km of main road network and 69 km of power lines in our main study area. In spring 2010, the construction and establishment of the two wind farms started in the hills along the spine road. To access the wind farms, 22 km of new roads were constructed, with 8.5 km of 36 kV power lines connecting to the existing power grid via the new utility station built in between the wind farms. At Jokkmokksliden, the full infrastructure was developed in 2010 and the first five wind turbines were constructed, with another five turbines being erected in 2011. At Storliden, the road network and power lines were established in 2010 and the turbines were put in place in 2011, starting to generate electricity in November 2011. The years 2010 and 2011 are hereafter referred to as the “construction” phase, as development continued throughout. In addition, to deliver the energy produced by the wind farm, the 30 kV power line stretching through the whole study area from east to the west was enlarged to a 145 kV power line during the 2010/2011 winter season when reindeer were absent from the area. All the Euclidian distances of the linear features, such as roads and power lines, were calculated using ArcGIS.
Detecting changes in use of movement corridors
To identify potential changes in the use of the area around the wind farm construction sites, we counted the number of crossings of the wind farm area and over the spine road by the GPS-tracked reindeer. Furthermore, we calculated the step length [distance between two GPS relocations (m/2h)] of the individual reindeer. We then tested the step-length before and during construction in a classical difference in difference (DID) setup (Angrist and Pischke 2009). The step length in the pre-construction phase was compared with the step length in the construction phase within 1, 2, 3, 4, 5 and 6 km buffer zones around the wind farm area (Sawyer et al. 2013). Thus, we estimated the effect close to the wind farm construction sites and used the surrounding area further away as control in two time periods. We used a standard two-sample t test to determine whether movement rates varied between the two periods and within which distance (buffer zone) a probable difference between pre-construction and construction no longer existed. The buffer distances were arbitrary chosen; however covering distances from the wind farm, which we know from the disturbance literature could have an effect on reindeer (Skarin and Åhman 2014).
Secondly, we used the Brownian Bridge Movement Model (BBMM) to estimate individual- and population-level movement corridors for GPS-collared reindeer for each study phase (Horne et al. 2007). Compared to the ordinary kernel home range estimation or minimum convex polygon (MCP), which both combine locations dependent on distance, the BBMM methodology provides a unique opportunity to identify reindeer movement and migration patterns as it is dependent on both time and distance between the locations (Horne et al. 2007; Sawyer et al. 2013). If a reindeer did not appear within 6 km of the wind farm area (Fig. 1) it was excluded from the movement corridor analysis during that season (Table 1). We estimated the utilisation distribution (UD) of an animal using a Brownian bridge kernel method (Calenge 2006; Horne et al. 2007). This method places a smoothing (kernel) function above each line connecting two successive GPS locations. The smoothing function is a combination of two bivariate normal density functions and a Brownian bridge density function connecting the two points (Horne et al. 2007). The R (R Core Team 2015) package “adehabitatHR” (Calenge 2006) was used for the analysis. The spatial extent of the UD was defined as the 99 % BBMM home range boundary and was displayed on a 50 m grid. We used an estimated location error of 20 m. The resolution chosen corresponded to the lowest resolution of the habitat variables used. Population-level movement corridors within each season were then estimated by averaging the individual UDs within each study phase (Sawyer et al. 2013). The UD at both individual and population level provides a probabilistic measure of the movement corridors, whilst the height of the UD reflects intensity of use and the contours of the UD represent the surface area of the corridors.
The change in the intensity of use was evaluated by determining change in use of “stopover” habitats (Sawyer et al. 2009) or high-use areas within the development area, as we treated the step length with a DID setup. Stopover habitats were defined as the highest 25 % quartile in the UD. At the individual level, we calculated the area of stopover habitat for each reindeer within each study phase within the 1, 2, 3, 4, 5 and 6 km buffer zones. We used a standard two-sample t test to determine whether there was a difference in the area amount of stopover habitats between pre-construction and construction phases for each of the buffer zones.
Habitat selection
RSFs provide a tool to estimate animal preference for, or avoidance of, certain habitats and linear structures such as roads and power lines, at multiple scales (Johnson 1980; Manly et al. 2002; Polfus et al. 2011). We developed RSF-models with a use-availability design (Manly et al. 2002) to evaluate whether the wind farm construction phase affected the reindeer habitat selection. To assess the habitat selection at landscape scale and the home range scale we developed the RSF-models within Johnson’s (1980) second-order (landscape, home range placement) selection and third-order (within home range) selection, respectively. We compared habitat variables at reindeer GPS locations to random available locations within the calving/post-calving range defined by the MCPs (computed from 95 % of all GPS-positions to exclude outliers), to assess the second-order selection. To assess the third-order selection we compared habitat variables at reindeer GPS locations to random available locations within the individual BBMM home ranges. Only data from collars that, at some point, were recorded within the 2 km buffer zone around the wind farm were included in the RSF-analysis since it was the change in use around the wind farm we wanted to investigate. To assess if there was avoidance of the wind farm area during the construction phase, the locations were split between pre-construction and construction phases, allowing an interaction between GPS-locations distances to the wind farm within each phase to be incorporated into the model. We generated available points using a 1:1 ratio of used to available locations. Habitat selection was evaluated using generalized linear mixed-models with each individual animal per year as a random factor to account for sample size and autocorrelation (Gillies et al. 2006). The R-package “lme4” (Bates et al. 2014) was used for the analysis. AIC-values were used to identify the most parsimonious model in the selection of prediction parameters. To validate the models with the best fit, we used a k-fold cross validation (Boyce et al. 2002). A training set of data was extracted and used to calculate a Spearman’s rank correlation between the observed positions therein and the predicted probability of selecting those positions using the model. The predicted probability was arbitrarily divided in ten equal bins. A testing ratio of 20 % was determined and a k-fold partition of five groups was used. This resulted in five correlations to evaluate the model fit.