Date of Award

Fall 12-16-2022

Level of Access Assigned by Author

Campus-Only Thesis

Degree Name

Master of Science (MS)


Ecology and Environmental Sciences


Cynthia Loftin

Second Committee Member

Daniel Hayes

Third Committee Member

Kasey Legaard

Additional Committee Members

Alyson McKnight

Roy M. Turner

Linda Welch


Biologists require accurate population estimates to effectively manage for threats to birds in a changing climate. However, traditional avian survey methods have limitations due to observer errors and visibility bias. Planes equipped with cameras for high-resolution photography offer solutions to the limitations of traditional surveys while minimizing disturbance created during ground surveys. Additionally, remotely sensed imagery provides a snapshot of populations, allowing for retroactive counts and assessments. These advantages can be prohibitively expensive, however, due to the labor required to acquire the imagery and manually interpret images. The development of automated processing tools for object detection and classification with Artificial Intelligence (AI) offers promise as an alternative, though there are few studies that examine how context affects the accuracy of automated population estimates. We developed an AI algorithm to automate the detection and species identification of ~37,000 colonial birds in plane-based imagery of nesting islands in the Gulf of Maine. We implemented the You Only Look Once model to automate image interpretation across landscape contexts and species, and we trained the algorithm with a subset of manual interpretations. The model performed variably across species and behavior with F1 scores ranging from 0.01 to 0.85. Objects of similar color to our focal species’ plumage were falsely detected as birds, and birds that were missed by the AI often lacked contrast with their background, were difficult to interpret by manual observers, or were underrepresented in the training data. We used gradient-boosted regression trees to demonstrate that species characteristics, animal behavior, land cover spectral class, and interpretability affected the algorithm’s ability to classify a bird to the correct species. Our case study exemplified the utility of closely investigating contributors of uncertainty in AI performance, and we framed our results in an ecological context to provide recommendations for addressing uncertainty in automated wildlife monitoring.

Available for download on Friday, January 31, 2025