Expanding a Chinook scale image archive for AI

Scales are aged manually, which is time consuming, increases variability in age estimates, and creates a lag in data availability. Using artificial intelligence (AI) to estimate age may help reduce processing time while standardizing age estimation and increasing sample throughput and data quality. Consequently, we propose to create a large archive of images of scales to train an AI to estimate Chinook salmon age. A team of international, federal, and hatchery personnel have developed an AI that estimates age of chum salmon scales with ~90% agreement, and this AI will be adapted to age Chinook salmon scales. Training, testing, and validation of the AI requires analysis of scale images representative of Chinook populations and age classes. We plan to expand upon the existing ADF&G scale image archive by adding recent images of scales from five Arctic-Yukon-Kuskokwim (AYK) Chinook stocks (~51,000 images; three scales each from ~17,000 individuals) to ensure stocks and age classes are better represented. Currently, the archive does not provide an adequate diversity of samples.



Chinook salmon ages estimated from scales by trained readers are used in fisheries management to estimate returns, examine productivity, and develop escapement goals. In recent years, size and age at maturity of Chinook have declined, and these changes are reflected in scale growth patterns. These changes, however, are not included in ADF&G’s image archive. By adding images of these more recent growth patterns, the AI will learn from as many types of scale patterns as possible.



Additional benefits of maintaining an image archive include retaining reader knowledge, availability for use in future research projects such as stock separation analysis, and reduction of physical storage issues. Once the AI is completed, it will be publicly available allowing entities to obtain age estimates without the need for extensive technical training.

Development and use of AI tools would significantly increase processing speeds and sample sizes and improve consistency of age estimates. A future benefit of AI could be to develop a tool to use scale pattern analysis for fine scale stock identification and genetic stock validation. AI-based data could better inform these models to increase their precision and accuracy. Current genetic technology does not allow for fine-scale geographic resolution as required by current management practices. AI-based scale pattern analysis could possibly be used to differentiate among groups on a finer scale than is currently possible with genetics.

Objectives:

1. Create a scale image archive representing all age classes and primary stocks present in the AYK region to aid in the development of an AI model for estimating Chinook salmon age.

2. Assist with adapting and evaluating an existing AI model to age Chinook salmon.