Determining salmon provenance with automated otolith reading

AbstractSynthetic otolith marks are used at hundreds of hatcheries throughout the Pacific Rim to record the release location of salmon. Each year, human readers examine tens of thousands of otolith samples to identify the marks in salmon that are caught. The data inform dynamic management practices that maximize allowable catch while preserving populations, and guide hatchery investments. However, the method is limited by the time required to process otoliths, the inability to distinguish between wild and un-marked hatchery fish, and in some cases classification processes are limited by the subjective decisions of human readers. Automated otolith reading using computer vision has the potential to improve on all three of these limitations. Our work advances the field of automated otolith reading through a novel otolith classification algorithm that uses two neural networks trained with an adversarial algorithm to achieve 93% classification accuracy between four hatchery marks and unmarked otoliths. The algorithm relies on hemisection images of the otolith exclusively: no additional biological data are needed. Our work demonstrates a novel technique with modest training requirements that achieves unprecedented accuracy. The method can be easily adopted in existing otolith labs, scaled to accommodate additional marks, and does not require tracking additional information about the fish that the otolith was retrieved from.Author summaryMany fish harvested in commercial fisheries have a bone-like structure in their inner ear called an otolith. Otoliths are useful because as they grow changes in the fish’s environment create unique patterns that can be interpreted years later—similar to the way tree rings can be used to determine a tree’s age and rate of growth in different seasons. Hatcheries use this phenomenon to create unique patterns in the otoliths of fish they release that identify their origin. Trained otolith “readers” examine tens of thousands of otolith samples as part of commercial fishery management every year. The data help resource managers protect returns to wild streams and help hatcheries track the survival rates of their releases. However, human reading is time consuming, limited in the types of patterns that can be identified and can be subjective. Applying modern machine learning techniques to otolith reading has the potential to improve on all of these challenges. We developed a computational method that identifies hatchery marks in digital images of salmon otoliths without human input. In a test with otoliths containing one of four hatchery marks and unmarked otoliths, the method accurately classified 93% of otoliths. The method can be adopted in otolith labs and can be further improved by expanding the sample size used to train and test the method.