Occupancy modeling of habitat use by white-tailed deer after more than a decade of exclusion in the boreal forest
Baribeau, Aimie; Tremblay, Jean-Pierre; Côté, Steeve (2022), Occupancy modeling of habitat use by white-tailed deer after more than a decade of exclusion in the boreal forest, Dryad, Dataset, https://doi.org/10.5061/dryad.1c59zw3zf
The exclusion of herbivores in forest areas is a strategy used to reduce the impact of selective browsing and increase the regeneration of desired plant species. On Anticosti Island (Québec, Canada), selective browsing by white-tailed deer prevents the regeneration of balsam fir – white birch forests leading to their conversion into white spruce forests. Large deer exclosures were established for ca . 10 to 12 years in clear-cuts with patches of residual forest from 2001 to 2006 to assist in the natural regeneration of fir stands and to provide shelter and food resources for deer. Our objective was to assess how deer use exclosures after the removal of fences according to their spatial configuration and habitat composition. We randomly distributed automatic cameras for periods of 14 days during summer in six exclosures ranging from 3.1 to 11.2 km2 (n=25 cameras per exclosure) from which deer were reduced for 10 to 12 years. We compared candidate occupancy models that included spatial configuration and food resource variables while simultaneously controlling for variables affecting detection probability. We obtained weak evidence that deer habitat use increased by 19% when forage resources, represented by the cover of Cornus canadensis, increased from 0 to 100%. None of the other variables (distance between the border of exclosures and cameras and distance between forest patches and cameras) was retained, suggesting that the use of regenerating forests by deer in summer after a period of exclusion is related to forage availability and therefore, any forest management that improves food production during summer should help maintain or increase habitat use by deer.
During the summer of 2020 (25 June – 16 August), we distributed 25 cameras (Bushnell Trophy, Spypoint Force 10, Reconyx PC800 and Spypoint Force 20) in each of the six exclosures according to a stratified random distribution in 3 different habitat types (regenerating cuts, residual mature forest without balsam fir [representing forest cover] and residual mature forest with balsam fir [representing forest cover and food resources]).
We deployed each camera as accurately as possible to the random locations. We mounted cameras at level, 75 cm from the ground on trees with a diameter at breast height of 8 to 10 cm and facing the North quadrant. We removed vegetation obstructing the camera’s lens. Cameras recorded pictures for 14 days in each exclosure, taking a systematic picture every 5 minutes for a total of 288 pictures per day.
To classify automatically the presence and absence of white-tailed deer, we analyzed 485,338 pictures collected during 2,100 camera trap days using an artificial intelligence from Tan and Le (2019). We classified the pictures with the neural network/architecture EfficientNet from the EfficientNet-Pytorch library (Melas-Kyriazi 2020). We used a sample of 2968 pictures pre-classified as deer presence (n=1484) and absence (n=1484) to train the model and 636 pictures to validate the model efficiency. The model classified the pictures with a degree of precision of 94%.
To estimate visibility and model the detection probability of each camera, we used a profile board composed of four panels of 50 x 25 cm to estimate the lateral cover in front of the camera. We took a measure at 5 and 10 meters in front of each camera, and then estimated the percentage of visibility at each site, following the approach in Ferron et al. (1996). The approach consist of averaging the grade of lateral cover of four squares on a blanket held vertically in front of the evaluator (1: 0-19% of visibility, 2: 20-39%, 3: 40-59%, 4: 60-79%, 5: 80-100%), then multiplying it by 20 and deducing 10, for example: ((1 + 3 + 4 + 4) / 4) * 20 – 10 = 50 % of lateral cover visibility at 5 m in front of the camera. The same approach was used at 10 m. We also recorded whether the camera was pointing at a deer trail or not.
At each camera trap site, we surveyed all saplings and merchant trees present per species in a circular area of 100 m2 (5.64 m radius) centered on the camera. We tallied saplings per species and diameter at breast height (DBH, 1.3 m) classes to compute basal area per species.
We measured vegetation of the ground layer in a 4 m2 circular plot located at 2 m in front of each camera. We estimated the percentage of cover of common herbaceous plants regularly browsed by deer on the island: Cornus canadensis, Maianthemum canadense, Chamaenerion angustifolium, Clintonia borealis, Rubus sp., and Poaceae sp. We kept only two species for the analysis (i.e., C. canadensis and Rubus spp.) because the global presence of the other species at all camera trap sites was less than 4%, and therefore their availability was not high enough to include these species in the analysis.
To define spatial configuration, we used ArcMap tools (ArcGIS v10.4; Environmental Systems Research Institute, Redland, CA, USA), to measure the distance of the closest forest cover and the shortest distance to the exclosure border (m) for each camera trap site. When the camera was located under forest cover, we set the distance at 0.
To obtain convergence, we scaled all the covariates (MacKenzie et al. 2018). Pairwise correlations between all variables were all below 0.48.
Data files require Excel to be open
Natural Sciences and Engineering Research Council of Canada, Award: RDCPJ525308-18