Machine learning applied to otolith microchemical data to discriminate stock of origin in salmon

Abstract

A key challenge in ecosystem-based fisheries management is sustainably harvesting co-occurring species and populations that differ in their vulnerability to fishing. This challenge is exemplified in western Alaska Chinook salmon, where recent population declines have led to closures of in-river subsistence fisheries, but identifying the cause of these declines is limited by an inability to identify the river of origin for marine-caught fish in offshore fisheries that target more abundant species and populations. This problem is particularly acute for estimating the impacts of bycatch on individual salmon populations which have demographic independence but no genetic differentiation from which stock assignments can be made. Here we use machine learning approaches to assess the efficacy of using the microchemical history preserved in fish otoliths to assign individuals to their river of origin in western Alaska. We tested the classification ability of three machine learning algorithms (Random Forest, K-Nearest Neighbors, and Support Vector Machines) combined with two time series smoothing techniques (Moving Average and Generalized Additive Models) to classify Chinook salmon to their river of origin using otolith time series data. Model accuracy ranged from 71.9% to 92.5%, with optimal performance achieved by Random Forest applied to GAM-smoothed data. Watershed-specific performance ranged from 86.8% to 93.7%, with most misclassifications occurring between spatially proximate Kuskokwim and Nushagak rivers. Raw predicted probabilities from classification algorithms were calibrated to reflect true class probabilities, enabling the incorporation of model results into decision analyses with explicit consideration of misclassification risk tolerance. The success of these models offers immediate utility for estimating marine mortality impacts across the regions three major river systems as well as an opportunity to understand commercial fisheries impacts on individual populations at substantially finer spatial scales than had been previously possible.