Deep Learning to Uncover Watershed Specific Characteristics in Salmon Otolith Patterns to Aid in Management of Pacific Salmon

Project Lead: Ben Makhlouf and Daniel Schindler, UW Aquatic and Fisheries Sciences

Data Science Lead: Valentina Staneva

Pacific salmon are vital to Alaskan communities, serving as a key cultural, subsistence, and economic resource. However, Chinook salmon populations in Western Alaska have declined sharply in recent years, particularly in the Yukon, Kuskokwim, and Nushagak River basins. While many hypotheses suggest marine mortality as a key factor, identifying the specific origins of fish caught in the Bering Sea has been challenging due to the limited genetic differentiation among these river systems.

Isotopic markers in otoliths (ear stones) of pacific salmon offer a promising alternative method as the ratio of strontium isotopes (Sr87/Sr86) in otoliths reflects the water chemistry of the rivers where fish originated. These metabolically inert structures grow sequentially over the lifetime of the individual, continually recording a sequential record of experienced isotopic values and allowing researchers to reconstruct life histories by matching these values to modeled ratios across large river basins. This method has been used to determine the origin of pacific salmon by isolating the portion of this time-series related to the juvenile freshwater lifestage in fish from a known river basin. However, because of overlapping isotope values among the three basins, single-point isotope measurements often fail to reliably assign fish to their natal rivers if their river basin is unknown, such is the case for fish caught in the open ocean. 

To overcome this limitation, we are applying time series clustering, machine learning, and otolith shape analysis to the full timeseries of isotopic data from birth until marine entry in known-origin Chinook salmon. By analyzing the full sequence of isotopic data recorded in the otolith over a fish’s lifetime, rather than a single value, our model can already distinguish the river of origin for more than 75% of individuals. Additionally, otolith shape analysis shows promise in further refining identification, particularly for at least one of the river systems. We have investigated and are beginning to more thoroughly understand which portions of the timeseries and features in otolith morphology may contribute towards reliable classification, which may give insight into how these tools may be used for similar applications in other systems. In addition, we have been able to begin to quality preprocessing and laboratory preparation techniques which may contribute towards model performance. This project has not only advanced my understanding of how data science can be applied to ecological research, but has also demonstrated the potential to extract valuable new insights from existing archival data using new tools and approaches.