Using Machine Learning to Support Salmon Population Monitoring
Computer vision, deep learning, and other artificial intelligence programs are revolutionizing the way our society processes and analyses large quantities of data. These tools, which have been pioneered over the last 40 years by computer science and technology industry leaders, now touch almost every part of our every day lives, from ordering a taxi, to restocking grocery store shelves. But far too often the benefits of these cutting-edge computing tools have not reached rural and remote communities in Canada.
Salmon population monitoring is an area where computer vision models have the potential to catalyze transformations that benefit both wild salmon and the people that depend on them. Rapid advances in video and data storage technology have led to the widespread use of underwater videography and sonar in salmon population monitoring. However, these continuously recorded datasets often span several months during the salmon migration season, and processing video and sonar outputs to make reliable counts typically requires large commitments of staff time and is often done after the field season limiting the ability of this information to inform in-season decisions.
Through a grant provided by the British Columbia Salmon Restoration and Innovation Fund, we are working with Simon Fraser University, the Heiltsuk, Kitasoo Xai’xais, and Haida Nations, Fisheries and Oceans Canada, and the Wild Salmon Center to develop computer vision models that automate salmon identification and counting from video and sonar monitoring projects. This technology will save vital staff time enabling broader application of video and sonar monitoring at lower cost. In addition, rapid video and sonar processing will provide vital in-season information on salmon returns, enabling precautionary management that reduces fishing impacts when runs are low, and mobilizes fishing opportunities when fish are abundant.
By the end of our three year project, we will produce several key products that can be used by salmon monitoring practitioners throughout BC and beyond:
- Open-source software for computer vision model that automates species identification and counting;
- Training dataset and open-source repository with >5000 observations of each species, enabling continuous learning and adaptation by computer vision models;
- Documentation and instructions enabling non-computer scientists to use software and validate results with their video or sonar datasets;
- A tested and transferable package of tools – video systems, power systems, satellite internet, and data processing – to enable widespread application of automated video monitoring.
Wild salmon face a rapidly changing climate and a future with no historical analog. They are incredibly resilient and are likely to remain capable of supporting harvest opportunities in many years. However, improved in-season monitoring will be an essential tool for government, First Nations, and local communities to understand unpredictable booms and busts in their local salmon stocks, and manage accordingly. By building these tools we are leveraging technology that has been decades in the making, and applying computer vision models to one of the most critical challenges in BC, monitoring and protecting wild salmon for future generations.