Snow sinking depth and forest canopy drive winter resource selection more than supplemental feeding in an alpine population of roe deer

Abstract

In alpine environments, snow typically reduces the accessibility of herbivores to food during winter and may hamper survival in those species with poor adaptation to move in deep snow. Supplemental feeding systems compensate for food limitation, but modify resource distribution and potentially affect individual space use. We investigated the importance of snow cover and supplemental feeding in shaping winter habitat use and selection of the European roe deer (Capreolus capreolus), a small deer species not specifically adapted to snow. We applied a used/available experimental design to assess the effects of snow cover on roe deer distribution at a fine scale and compared this approach with remotely sensed satellite data, available at moderate spatial resolution (snow MODIS). Based on this, we developed a resource selection function. We found a strong selection for habitat spots covered by forest where snow sinking depth was less pronounced, likely providing thermal and hiding protection on the one side and minimising the effect of snow on locomotion on the other. Roe deer showed only a minor preference for sites in proximity to feeding stations, possibly compensating the costs of access to these sites by means of a ‘trail-making’ behaviour. Snow cover assessed by moderate resolution satellite was not proportional to roe deer probability of use, highlighting the importance of local information on snow quality and distribution to complement remotely sensed data.

Appendices

Appendix 1

Tool for empirical assessment of snow sinking depth

The battage probe (Fig. S1) is a percussional tool made of tubular elements marked with a centimeter scale. A driving pole is put above these elements, whose number depends on the overall thickness of the snow layer. An additional weight with a central hole is placed on the top of the driving pole and released.

Tests of the hardness of snow layers are performed as follows:

• The probe without any additional weight (i.e. tubolar elements + driving pole) is placed on the snow surface, and its sinking is measured.

• The additional weight is added to the whole tool, without releasing it (i.e. at the bottom of the driving pole); the sinking depth is measured.

• The additional weight is released from points at increased height on the driving pole and this operation is repeated a certain number of times for each height. The procedure goes on until the probe has entirely entered into the snow layer. The hardness of the layer is defined as

  $$ R = P*n*H*{D}^{-1} + P + A $$

  where P = additional weight (N); n = number of times the weight is dropped at a certain height H; H = dropping height (m); D = sinking height of the probe (m); and A = weight of the tubolar elements (N). This operation is repeated each time the snow probe encounters a new layer, until the ground.

In our experiment, we calibrated this tool to evaluate the pressure that a roe deer exerts on the snow surface and consequently to get a realistic estimate of its sinking in the snow. We provided the tubular elements of two further additional weights of mass equal to 2.5 kg each (Fig. S2a). Then, the overall mass of the probe was equal to 7 kg (2.5 * 2 + 2 kg of the tubular elements and the additional weight provided with the probe). The resulting weight of 70 N approximates the force exerted by the leg of a 20 kg roe deer during a walking mode, under the assumption that roe deer walking motion can be simplified by a static assessment and that at each step the pressure is equally distributed on all the three legs in contact with the ground. Moreover, we substituted the steel tip of the probe with the distal part of a roe deer hindleg (Fig. S2b) to mimic more accurately the impact of roe deer on the snow. The snow sinking measurements were performed by gently releasing the probe in the snow and measuring the reached depth with a rigid meter, without assessing the specific hardness of the snow layer.

Appendix 2

Procedure to identify the initial full model and for model selection

We identified a set of potentially biologically meaningful covariates to analyse roe deer winter resource selection and specifically: a spline of the week to take into account the time autocorrelation; the distance from the closest feeding station; remote index of snow cover presence derived from MODIS; empirically recorded index of snow cover patchiness; average snow sinking depth; average snow thickness; canopy presence, terrain slope and solar radiation; individual sex; and identity of the individual.

Based on the collinearity analysis (Table 2), we retained average snow sinking depth but not average snow thickness in the full model. Moreover, we also fitted a two-way interaction between canopy presence and terrain slope.

Thus, the final set of covariates to fit in a model included a spline of the week; the distance from the closest feeding station; snow cover presence derived from MODIS; average snow sinking depth; index of snow cover patchiness; a two-way interaction between canopy presence and slope; solar radiation; sex; and the identity of the individual fitted as random effect.

The procedure of model selection we used involved several steps. First, we assessed the importance of the contribution of the random effect of the individual to determine the goodness-of-fit of the initial full model (Table 4). We fitted a GLMM including the terms above mentioned (model random, AIC = 312.9), as well as a GLM with the same framework but not the random effect (model fixed, AIC = 310.86). We computed the percentage of variation explained by the random effect of the individual as the ratio between (i) the difference between the deviance of the model with random effect and the model without it and (ii) the deviance of the null model. We found that this value was very low (3.15e-13); therefore, we decided to remove the random effect, to perform the rest of the analyses using generalised linear models.

We used an AIC-based model selection (Burnham and Anderson 2002) to determine the models which better explained the variation of the response variable (∆AIC < 2). We thus retained 13 models (Table 5), neither of which included the spline of the week or the index of snow cover. We computed the predictors’ weight on the retained models. We then proceeded with two parallel approaches to obtain the final models. First, in compliance with the multi-model inference theoretical framework, we performed model averaging on these models. We obtained a final averaged model with weighted coefficients and standard errors (Table 6). Moreover, in compliance with the principle of parsimony, we obtained a simplified version of the model by retaining only those covariates included in all models with ∆AIC < 2, i.e. with predictor’s weight = 1. Thus, this version of the final model included the canopy presence, the distance from the closest feeding station and the average snow sinking depth (Table 6). The estimated model was validated according to the classic Resource Selection Analysis framework (Boyce et al. 2002).