Incorporating habitat selection does not account for nonrandom camera deployment in a design-based viewshed density estimator

Camera trap-based abundance estimators are increasingly used for population size estimation in the absence of marked individuals. One approach is to relate animal detections to the space sampled by each cameras viewable area, which results in viewshed density estimates that can be extrapolated to the broader sampling area to obtain abundance. The assumption is that spatial variation in local abundance corresponds to the collective viewsheds of cameras. Therefore, these design-based viewshed density estimators require that camera locations be representative of the study area. This assumption can be met with spatially balanced probability sampling, such as a generalized random tessellation stratified design. However, the random placement of cameras can be restrictive in practice. Using simulations and an empirical study, we evaluated an extension of the instantaneous sampling estimator to account for unmodeled spatial variation in local abundance using independent predictions of relative habitat use from GPS telemetry data. We applied this approach to a fenced population of white-tailed deer (Odocoileus virginianus), some of which were GPS collared, where timelapse photography data were collected using random and nonrandom camera placements simultaneously during two consecutive winter seasons. Our simulations showed that this approach neither reduces bias nor improves the precision of abundance estimates considerably. Specifically, we found little support that calibrating estimates based on habitat selection analysis would produce unbiased results when sampling is spatially unbalanced and focused on areas of high animal use instead of being representative of the study area. Our findings underscore the need for randomized sampling when the goal is to provide unbiased population size estimates for unmarked wildlife populations using design-based viewshed density estimators. Attempts to relax the model assumptions must be both theoretically sound and practically feasible, otherwise there would be a risk of misleading management decisions by using unreliable estimates of population size.