A collaborative and near-comprehensive North Pacific humpback whale photo-ID dataset

We present an ocean-basin-scale dataset that includes tail fluke photographic identification (photo-ID) and encounter data for most living individual humpback whales (Megaptera novaeangliae) in the North Pacific Ocean. The dataset was built through a broad collaboration combining 39 separate curated photo-ID catalogs, supplemented with community science data. Data from throughout the North Pacific were aggregated into 13 regions, including six breeding regions, six feeding regions, and one migratory corridor. All images were compared with minimal pre-processing using a recently developed image recognition algorithm based on machine learning through artificial intelligence; this system is capable of rapidly detecting matches between individuals with an estimated 97-99% accuracy. For the 2001-2021 study period, a total of 27,956 unique individuals were documented in 157,350 encounters. Each individual was encountered, on average, in 5.6 sampling periods (i.e., breeding and feeding seasons), with an annual average of 87% of whales encountered in more than one season. The combined dataset and image recognition tool represents a living and accessible resource for collaborative, basin-wide studies of a keystone marine mammal in a time of rapid ecological change.

Quantifying the age structure of free-ranging delphinid populations: Testing the accuracy of Unoccupied Aerial System photogrammetry

Understanding the population health status of long-lived and slow-reproducing species is critical for their management. However, it can take decades with traditional monitoring techniques to detect population-level changes in demographic parameters. Early detection of the effects of environmental and anthropogenic stressors on vital rates would aid in forecasting changes in population dynamics and therefore inform management efforts. Changes in vital rates strongly correlate with deviations in population growth, highlighting the need for novel approaches that can provide early warning signs of population decline (e.g., changes in age structure). We tested a novel and frequentist approach, using Unoccupied Aerial System (UAS) photogrammetry, to assess the population age structure of small delphinids. First, we measured the precision and accuracy of UAS photogrammetry in estimating total body length (TL) of trained bottlenose dolphins (Tursiops truncatus). Using a log-transformed linear model, we estimated TL using the blowhole to dorsal fin distance (BHDF) for surfacing animals. To test the performance of UAS photogrammetry to age-classify individuals, we then used length measurements from a 35-year dataset from a free-ranging bottlenose dolphin community to simulate UAS estimates of BHDF and TL. We tested five age classifiers and determined where young individuals (<10 years) were assigned when misclassified. Finally, we tested whether UAS-simulated BHDF only or the associated TL estimates provided better classifications. TL of surfacing dolphins was overestimated by 3.3% ±3.1% based on UAS-estimated BHDF. Our age classifiers performed best in predicting age-class when using broader and fewer (two and three) age-class bins with ~80% and ~72% assignment performance, respectively. Overall, 72.5%-93% of the individuals were correctly classified within 2 years of their actual age-class bin. Similar classification performances were obtained using both proxies. UAS photogrammetry is a non-invasive, inexpensive, and effective method to estimate TL and age-class of free-swimming dolphins. UAS photogrammetry can facilitate the detection of early signs of population changes, which can provide important insights for timely management decisions.

Optimizing automated photo identification for population assessments

Several legal acts mandate that management agencies regularly assess biological populations. For species with distinct markings, these assessments can be conducted noninvasively via capture-recapture and photographic identification (photo-ID), which involves processing considerable quantities of photographic data. To ease this burden, agencies increasingly rely on automated identification (ID) algorithms. Identification algorithms present agencies with an opportunity-reducing the cost of population assessments-and a challenge-propagating misidentifications into abundance estimates at a large scale. We explored several strategies for generating capture histories with an ID algorithm, evaluating trade-offs between labor costs and estimation error in a hypothetical population assessment. To that end, we conducted a simulation study informed by 39 photo-ID datasets representing 24 cetacean species. We fed the results into a custom optimization tool to discern the optimal strategy for each dataset. Our strategies included choosing between truly and partially automated photo-ID and, in the case of the latter, choosing the number of suggested matches to inspect. True automation was optimal for datasets for which the algorithm identified individuals well. As identification performance declined, the optimization recommended that users inspect more suggested matches from the ID algorithm, particularly for small datasets. False negatives (i.e., individual was resighted but erroneously marked as a first capture) strongly predicted estimation error. A 2% increase in the false negative rate translated to a 5% increase in the relative bias in abundance estimates. Our framework can be used to estimate expected error of the abundance estimate, project labor effort, and find the optimal strategy for a dataset and algorithm. We recommend estimating a strategy's false negative rate before implementing the strategy in a population assessment. Our framework provides organizations with insights into the conservation benefits and consequences of automation as conservation enters a new era of artificial intelligence for population assessments.