Vantage point photography and deep learning methods save time in monitoring seabird nesting colonies

Abstract Monitoring seabird colonies is essential for assessing population health and sustainability amid increasing marine industry and climate change. Advancements in photography have led to high-resolution photogrammetry techniques for monitoring seabird colonies. However, manual counting of birds and nests in images, potentially over multiple dates and seasons, is time-consuming and has limited a wider adoption of photogrammetry in colony monitoring. We addressed the task of automatically counting cormorants and their nests using an SLT camera, a Gigapan robotic camera mount, and image-stitching software. We applied this system to Vancouver’s Ironworkers Memorial Second Narrows Bridge, home to British Columbia’s largest Nannopterum auritum (Double-crested Cormorant) colony. The system takes overlapping images of the colony from a vantage point to create a panoramic image. We took 23 images of the bridge between April and September 2021. A subset of these images was used to train a deep-learning model that became the foundation of an automated pipeline to detect cormorants of different sizes, positions (standing, incubating, or sunning with wings outstretched), and nests (including only partial glimpses among the bridge girders). Our pipeline demonstrated potential for monitoring cormorant populations, by lowering manual effort while achieving high agreement with manual counts. Specifically, our pipeline reduced the manual time required to process images by 96% while achieving an average agreement of 93.6% between manual and automated counts for both cormorants and nests. We found reduced performance from an application of our model to images of a novel colony; however, we suggest that with additional model training and fine-tuning, our pipeline should provide an efficient and accurate alternative to manual counts for other colonial bird monitoring contexts. Our study showcases that high-resolution photogrammetry combined with deep learning methods enables the automatic identification and counting of birds and nests, significantly reducing the time and effort of long-term monitoring of colonially nesting birds.