Using Deep Image Segmentation Methods to Detect and Classify Salmon Species from ARIS Sonar Images

Improvement of species composition estimates in the Fraser River during sockeye salmon migration is a priority of the Fraser River Salmon Fisheries Management for the 2021 projects. Given the low returns of Fraser River sockeye salmon and poor catches from test fishing programs in recent years, species composition estimates based on the catch data have become increasingly unreliable for the in-season management of Fraser stocks. To address this challenge, alternatives methods must be developed to provide reliable species composition estimates for both the daily and the seasonal total sockeye abundance estimates.

For this proposed project, we will use image data from a newer generation of imaging sonar (ARIS), which offers a higher spatial resolution than the DIDSON sonar used in the previous projects. This project is aimed at delivering the following 1) improved accuracy and reliability in species classification; 2) automatic measurements of fish length and other feature variables with confidence rankings that can be used in the PSC mixture model for species composition estimation.