Effects of distance on detectability of Arctic waterfowl using double‐observer sampling during helicopter surveys

Aerial survey is an important, widely employed approach for estimating free‐ranging wildlife over large or inaccessible study areas. We studied how a distance covariate influenced probability of double‐observer detections for birds counted during a helicopter survey in Canada’s central Arctic. Two observers, one behind the other but visually obscured from each other, counted birds in an incompletely shared field of view to a distance of 200 m. Each observer assigned detections to one of five 40‐m distance bins, guided by semi‐transparent marks on aircraft windows. Detections were recorded with distance bin, taxonomic group, wing‐flapping behavior, and group size. We compared two general model‐based estimation approaches pertinent to sampling wildlife under such situations. One was based on double‐observer methods without distance information, that provide sampling analogous to that required for mark–recapture (MR) estimation of detection probability, [Formula: see text] , and group abundance, [Formula: see text] , along a fixed‐width strip transect. The other method incorporated double‐observer MR with a categorical distance covariate (MRD). A priori, we were concerned that estimators from MR models were compromised by heterogeneity in [Formula: see text] due to un‐modeled distance information; that is, more distant birds are less likely to be detected by both observers, with the predicted effect that [Formula: see text] would be biased high, and [Formula: see text] biased low. We found that, despite increased complexity, MRD models (ΔAICc range: 0–16) fit data far better than MR models (ΔAICc range: 204–258). However, contrary to expectation, the more naïve MR estimators of [Formula: see text] were biased low in all cases, but only by 2%–5% in most cases. We suspect that this apparently anomalous finding was the result of specific limitations to, and trade‐offs in, visibility by observers on the survey platform used. While MR models provided acceptable point estimates of group abundance, their far higher stranded errors (0%–40%) compared to MRD estimates would compromise ability to detect temporal or spatial differences in abundance. Given improved precision of MRD models relative to MR models, and the possibility of bias when using MR methods from other survey platforms, we recommend avian ecologists use MRD protocols and estimation procedures when surveying Arctic bird populations.