Between spring and fall each year in coastal British Columbia, when salmon migrate upstream, the region’s First Nations manually count the number of fish passing through to get a sense of how healthy the population is. But it’s work that takes place in remote and hard-to-access streams of the province, making it laborious, time-consuming, and often error-prone.
So for a recent study, marine scientists, computer scientists and conservation practitioners partnered with Indigenous-led fisheries organizations to build and deploy an automated system to monitor and count salmon.
This first-of-its-kind tool harnesses the power of artificial intelligence to “learn” how to differentiate objects using computer vision algorithms. It can recognize and count 12 species of fish found in the Pacific Northwest, including the five species of wild Pacific salmon, by merely scanning video clips. The study was published in the journal Frontiers in Marine Science.
“This is the first time that anyone has automated counting of salmon from a video,” said Will Atlas, a salmon watershed scientist at the Oregon-based Wild Salmon Center. “We’ve come sort of the closest to having a tool that’s ready to be rolled out into actual management applications.”
In coastal British Columbia, Pacific salmon holds a unique place as a culturally revered fish for Indigenous peoples, and is a prized delicacy for seafood aficionados. The many coastal First Nations had sustainably managed salmon numbers for thousands of years, until logging and overfishing destroyed the delicate balance in the last century. As a result, the number of salmon returning to the many creeks and rivers where they spawn has fluctuated dramatically, casting doubts about their future.
“A central part of managing and conserving salmon is monitoring the number of adult salmon that return to the river to spawn,” Atlas told Mongabay.
Doing so manually, however, just isn’t feasible. “It’s challenging work because we’re at the whim of Mother Nature and the environment,” said fisheries biologist Mark Cleveland from the Indigenous-led Gitanyow Fisheries Authority in Kitwanga, British Columbia.
To train their AI-based tool, researchers used more than half a million video clips recorded by the Gitanyow Fisheries Authority and the Skeena Fisheries Commission. In recent years, these two Indigenous-led fisheries management organizations have begun using high-definition underwater cameras to monitor salmon migration in the Kitwanga and Bear rivers. But they still depend on humans to review the video and count the salmon.
In its initial stages of development, the AI-based tool needed humans to “teach” it to identify salmon — a task field technicians helped with by annotating salmon in the video clips. Over time, the tool learned to recognize the fish so well that it got it right seven times out of 10. Its accuracy surpassed 90% for sockeye (Oncorhynchus nerka) and coho salmon (Oncorhynchus kisutch), two of the important species in the North Pacific.
However, its accuracy was low in identifying pink (Oncorhynchus gorbuscha) and chinook salmon (Oncorhynchus tshawytscha) becUSause the individuals of these species differ in their looks. During the spawning season, the male pink salmon develops a large hump and hooked jaw, and the chinook salmon change colors.