|
Abstract:
|
The use of unmanned aerial vehicles is becoming increasingly popular to survey animal populations. However, the benefits of collecting more data, faster, may not be fully exploited under the traditional analytical framework. We developed a method that leverages aerial imagery data for population modeling using a record linkage method and deduplication. Deduplicating individuals in image overlaps with distortion requires realigning observed point patterns optimally; however, popular machine learning algorithms for image stitching do not often account for alignment uncertainty. Moreover, duplicated individuals can provide insight about detection probability when overlaps are viewed as replicated surveys. Our model resolves individual identities by “linking” observed locations to latent activity centers, and estimates detection probability and total population as informed by the linkage structure. We developed a unified hierarchical deduplication model to avoid single-direction error propagation that is common in two-stage models. We illustrate our method through a case study of aerial survey images of sea otters in Glacier Bay, Alaska.
|
