Applying computer vision to accelerate monitoring and analysis of bird incubation behaviors: a case study using common eider nest camera footage

Abstract Advances in camera and data storage technology have revolutionized the ability of scientists to acquire large volumes of finely resolved wildlife monitoring data. This is especially valuable for breeding bird research, which often requires and benefits from continuous nest monitoring, which may extend a month or more. Though high‐quality imagery may yield valuable insights, the sheer volume of data can create processing bottlenecks. Furthermore, achieving uniformity across projects and years is difficult given individual‐level differences in data processing by manual reviewers. To address this problem, we paired a custom trained You Only Look Once version 7 (YOLOv7) model with the StrongSORT tracking algorithm to analyze videos of nesting common eiders ( Somateria mollissima ) collected from barrier islands along the Beaufort Sea coast in Alaska. We used our computer vision pipeline to process footage three times faster than manual review while matching human observer accuracy in recording nest attendance and disturbances. To evaluate the effectiveness of our trained pipeline, we analyzed novel footage from a different year. The automated part of the pipeline performed well when birds were relatively large in the frame. However, performance declined for birds occupying a small frame area, which occurred when the camera was farther away from the nest and not zoomed. When birds are smaller in the frame, they are more susceptible to being obscured by rain or fog on the lens, as well as by other birds positioned in front of them. Additionally, detecting birds that occupy a small area of the frame can be more challenging in complex backgrounds, particularly under difficult lighting conditions, such as when the sun backlights the bird, or due to specific behaviors, like when birds hunker down to minimize their silhouette in response to perceived threats. To enhance performance, we recommend that researchers position cameras closer to nests whenever feasible or utilize zoom lenses. Importantly, our pipeline is designed to be species‐agnostic, allowing for easy adaptation to various nesting bird species.