Monitoring cetaceans with computer vision and hierarchical models

Abstract

Several legal acts mandate that wildlife management agencies regularly assess cetacean populations. These population assessments often consist of space-use and abundance estimates from satellite tag (telemetry) and photographic identification (photo-ID) data. This dissertation explores how to process these data more efficiently with advances in computer vision, and how to model these data more effectively with hierarchical modeling. In Chapter 2, I introduced a multi-species photo-ID algorithm that jointly predicts species and individual ID, allowing information sharing across species within a single neural network. I trained the algorithm on 50,796 images from 39 datasets representing 24 cetacean species. I tested the algorithm on 21,192 images across these same datasets, where it achieved a mean average precision of 0.869, with ten catalogs exceeding 0.95. The algorithm performed best on high-quality images of species identified with nicks and notches along the dorsal fin and performed worse with low-quality images of indistinct animals. What do these performance metrics mean in terms of abundance estimation? To what degree can we automate photo-ID while still producing reliable inference? I explored these questions with a simulation study, informed by the 39 datasets above. I found that the algorithm can reduce labor effort associated with photo-ID while minimally biasing abundance. Indeed, 22 of 39 evaluated datasets achieved minimal relative bias (less than 10%) in a low labor effort scenario. False negative rates strongly predicted abundance estimation error, with a 2% increase in false negatives translating to 5% increased relative bias. Finally, I demonstrated how to jointly estimate abundance, individual space-use, and stock boundaries by integrating photo-ID and telemetry data with spatial capture-recapture. Applied to rough-toothed dolphins around Kaua'i Island, this approach simultaneously estimated population size (1,571 marked individuals; 96% CI: 1,398-1,763) and defined habitat boundaries, which encompassed an approximately 8,000 km² around Kaua'i and Ni'ihau islands. Together, these chapters show that integrating computer vision with statistical modeling makes cetacean stock assessment more informative and efficient. Each chapter of the dissertation was built on open-source code and offered guidelines on best practices on effectively using these techniques.