New research shows artificial intelligence is better than any human at counting and identifying salmon underwater.
And with more training, it could be an invaluable tool for salmon fisheries managers and researchers.
Researchers from Oregon and BC worked together during the last two years to create AI-powered tracking and counting systems. They set them up at Indigenous-run weirs on the Kitwanga and Bear Rivers, part of the Skeena watershed, and at the Coquitlam Dam.
The systems were able to process recordings at a rate of one hour of video in about 60 seconds. They were very good at detecting sockeye and coho salmon, but had a difficult time correctly identifying pink and Chinook salmon.
“Pink salmon exhibit extreme sexual dimorphism, with males developing large humps and hooked jaws during spawning season,” researchers say in the study. “These differences between male and female morphology likely pose unique challenges for automated detection unless accounted for in future training procedures and analysis. Chinook salmon have a higher degree of variability in their coloration, age, size, and secondary sexual characteristics and likely require larger quantities of training data.”
Their prototype system can be replicated cheaply, for about $3,500 USD, with off-the-shelf technology. With an Internet connection via Starlink satellite to cloud storage, a remote counting system can collect and upload data from almost anywhere. Similar systems are already being used in remote locations on the BC coast by First Nations fisheries.
The research, published last month in Frontiers in Marine Science, shows that with specific training, AI has a lot of potential to save time and money, while providing accurate information about salmon stocks. The system could also be used to research other species.
