Have you ever wondered how salmonids migrating up a river get tracked? And why this matters?
Fish counts help biologists understand whether fish populations are healthy, whether conservation efforts are working, and how ecosystems change over time. Right now, this often means people are spending hours watching video from underwater cameras. The work is slow and expensive.
João Rodrigo da Silva Martins from the University of Iceland and The Marine and Freshwater Institute wanted to change this. Together with his supervisor Hafsteinn Einarsson and colleagues Hlynur Bárðarson and Jóhannes Guðbrandsson he explored whether AI could automate part of this work. The publication can be found here here.
Using AI to track fish
The team used modern AI to identify fish species automatically from underwater videos.
Around Iceland there are automated fish counting stations in rivers where underwater footage gets recorded. Experts watch thousands of these videos to classify the fish they see. João tested whether AI could take over some of this work. The results were promising. The AI could identify fish accurately and quickly. This would give experts more time to monitor rivers and focus on more interesting scientific questions.
On the right: a fish counter installed in rivers. On the left: an example of underwater footage of a fish.
The model identified whether a fish was present in a video frame with 99% accuracy. This is almost the same level of accuracy as a human expert.
The system could also tell the difference between salmon, trout, and charr with 97% accuracy. This is impressive because even experienced freshwater scientists can sometimes struggle to tell these species apart.
How does AI recognise a fish?
The researchers used an AI model that had already learned from millions of online images and captions. Because the model had seen so many examples, the researchers could ask questions such as: “Does this look like a salmon-like fish swimming?”
The method works in two steps. First, the detections stage filters the raw video. It removes frames without fish and finds clips that contain a single fish.
Second, the classification stage identifies the species. The model creates a unique “fingerprint” for each frame containing a fish. It then combines information from several frames and creates an average fingerprint. This helps reduce mistakes caused by poor lighting or unusual fish movements. The system then decides whether the fish is a salmon, trout, or charr.
This image helps researchers understand how the AI model identifies fish species. Red areas show the parts of the image that most affect the model’s performance when hidden. This suggests that the model relies on fish body shape and head features rather than background information.
Why this research matters
This method could help fisheries management agencies, such as The Marine and Freshwater Institution here in Iceland. But it is also useful for similar organisations abroad.
Many organizations collect more fish videos than people can realistically review by hand. As a result, a lot of valuable footage is never used.
AI could act as a first filter. It could identify clips that contain fish and classify species that are easy to recognize. Experts would then only need to review unusual fish or difficult cases.
Better and faster fish counts could improve stock assessments, support more accurate fishing quotas, and lead to better conservation decisions.
João and his colleagues want to make AI tools easier to use in ecology and environmental science. These fields often have large amounts of data but too few people to process it. Their work shows that AI can help solve this problem. The hope is that this approach will help research groups, agencies, and non-profit organisations make better use of their data.
A salmon swimming past a fish counter.
