Habitat selection by ice-associated pinnipeds near St. Lawrence Island, Alaska in March 2001

Abstract

Aerial surveys of ice-associated pinnipeds were conducted south of St. Lawrence Island in March 2001. The observed distributions of bearded seals (Erignathus barbatus), ribbon seals (Phoca fasciata), ringed seals (P. hispida), spotted seals (P. largha), and walruses (Odobenus rosmarus) were compared to the distributions of ice habitat types and benthic communities. Randomization tests were used to investigate habitat selection for each species. Both ringed seals and walruses preferred large ice floes (>48 m in diameter) that were common in the interior ice pack. Spotted seals favored smaller ice floes (<20 m in diameter) common near the ice edge, and bearded seals avoided large floes and preferred transitional habitat between small and large floes. Ringed seals also seemed to prefer areas with greater than 90% sea ice coverage, and bearded seals preferred 70–90% sea ice coverage while avoiding areas with greater than 90% coverage. All species, except spotted seals, were seen most frequently in a region of high benthic biomass, and randomization tests suggested that bearded seals actively selected that region.

Appendix

Correlated-random simulations of expected habitat use values

Each species' selection of habitats was analyzed by comparing the observed distribution of the species among habitat types (ice-coverage categories, floe types, and benthic regions) to the expected random distribution with respect to habitat types. The expected distribution of each species was characterized via repeated simulations. These simulations included spatial autocorrelation among observations to address the concern that the distribution of animals was likely to be patchy or clustered (i.e., a correlated-random distribution). The spatial autocorrelation between observations was assumed to be inversely related to distance, as predicted by an inverse-distance weighted correlation (IDW) model. The relationship between correlation and distance was controlled by the power (p) to which the inverse-distance was raised:

$${Y_{i} = {\sum\limits_{j \ne i} {{\left\{ {{\left( {{{d^{{ - p}}_{{ij}} } \over {{\sum\limits_{j \ne i} {d^{{ - p}}_{{ij}} } }}}} \right)} \cdot Y_{j} } \right\}}} } + \varepsilon ,}$$

(1)

where: Y _i = the value of Y at location i, d _ij = the distance between location i and j, p = the power to which the inverse-distance was raised, Y _j = the value of Y at location j, and ε = an independent, random error. This correlation model was only applied to data collected on the same survey day because data collected on separate days were assumed to be independent. Since the data were binary, the IDW model essentially predicted the probability that Y _i =1 given the values of, and distances to, all Y _j . The value of p for each species was chosen to minimize the difference (measured as mean squared error) between the observed values for each survey segment and the values predicted by an IDW model with power, p. For bearded seals, the power value was 2.14; for ribbon seals it was 0.65; for ringed seals it was 0.93; for spotted seals it was 0.79; and for walruses it was 1.21.

In the first stage of the correlated-random simulations, the observed presence-absence data were randomly reallocated to the survey segments without regard to spatial location, date, or habitat variables (i.e., simple permutation of the data among segments). The results from this stage were used as expected values for a random distribution and were used in randomization tests for comparison with analyses using the correlated-random simulations. In the second stage, the randomly allocated presence and absence values were iteratively reallocated based on predictions by the IDW model (Cressie 1991). During each step in the iterative process, a random pair of segments (a, b) from one survey day was chosen for which Y _a =1 and Y _b  =0. The values of Y _a and Y _b were then swapped (or not) based on the predicted probabilities from the IDW model:

$${{\rm{swap}}\,\,\,{\rm{value}} = {{\Pr {\left( {Y_{a} = 0} \right)} \cdot \Pr {\left( {Y_{b} = 1} \right)}} \over {\Pr {\left( {Y_{a} = 1} \right)} \cdot \Pr {\left( {Y_{b} = 0} \right)}}}}$$

(2)

The values of Y _a and Y _b were always swapped when swap value ≥1; otherwise, the values were swapped with probability = swap value. This iterative process of randomly selecting pairs and then swapping their values based on the IDW model was repeated as many times as there were unique (Y _a =1, Y _b =0) pairs for each survey day. The large number of repetitions ensured that the resulting distribution approximated a random distribution with spatial autocorrelation defined by the IDW model. Both stages of the simulation procedure (permutation followed by iterative swapping) were repeated 1,000 times for each habitat category (ice coverage, floe type, and benthic region), and segments with animals present in each habitat type were tallied each time. The resulting tallies were the expected values for correlated-random distributions of animals.