Observer coverage and interaction rates determine the choice between design-based and machine learning bycatch...

Robust estimates of the magnitude of protected species bycatch are essential for effective fisheries management and marine conservation. Existing bycatch estimation methods typically struggle to accurately and precisely estimate the bycatch of rare species as a result of the statistical assumptions necessary to overcome infrequent encounters and, often, low observer coverage. We compared the widely used Horvitz-Thompson design-based estimator with a machine learning framework for estimating protected species bycatch based on ensemble random forests. We simulated reduced observer coverage in the 100% observed Hawaii shallow-set pelagic longline fishery and compared the two estimation methods for five species: oceanic whitetip sharks (Carcharhinus longimanus), Laysan albatross (Phoebastria immutabilis), black-footed albatross (Phoebastria nigripes), loggerhead sea turtles (Caretta caretta), and leatherback sea turtles (Dermochelys coriacea). Machine learning-based estimates were more precise than design-based estimates for all species, but annual and long-term accuracy varied with bycatch rates and spatial clustering of interactions. The machine learning approach improved the precision of bycatch estimates for oceanic whitetip sharks (3.5% of sets with interaction) without reducing long-term accuracy. As interaction rates declined below this level, the ensemble random forests method introduced increasing bias and resulted in a preference for the design-based estimator in real-world applications. The ensemble random forests approach relaxed the trade-off between accurate, precise bycatch estimates and observer coverage for species with greater than 3% interaction rates. However, high observer coverage remains irreplaceable for effectively estimating bycatch for the rarest species, and the imprecision of design-based estimates should be accounted for in management decisions.