During Alaska stream research, teams developed an automated salmon-counting system to track fish in sonar video and reduce reliance on manual counting

We may think that any monitoring conducted by a government organization in 2026 would happen using drones, sensors, and some AI dashboard performing calculations behind the scenes. However, as far as salmon fisheries in Alaska go, the process is far from automatic. According to the Alaska Department of Fish and Game, a significant proportion of the salmon counting process carried out in Alaska is based on a method where a technician standing along the banks of a river clicks the counter every time he sees a salmon swim past; sometimes for ten minutes every hour, throughout the entire summer. This ten-minute-per-hour sampling routine is the method used at fish-counting towers, one of several manual techniques the department relies on. These are not just empty numbers. They are used directly to determine the number of fish that can be caught commercially and by recreational fishermen and also if enough salmon have survived for breeding.

The century-old jobs still keep Alaska's fisheries running

Department biologists also use several other rudimentary methods to obtain these figures: weirs, which involve the use of a fence to channel all the fish into one exit so that every single one is counted manually; towers, where technicians watch and tally fish for set portions of each hour; aerial and ground surveys conducted by skilled observers, either flown over or walked along the river; and mark and recapture studies, where fish are marked and spotted again after some time to determine their population.

The only way to count the wide, muddy-brown, glacier-fed rivers where nothing like this works is “sonar,” which sends out sound pulses and listens for the echoes bouncing off passing fish. It is a remarkable piece of ingenuity. But there is one aspect of this system that often goes unrecognized: even sonar images must be analyzed, frame by frame, by humans watching video footage to separate fish from floating debris.

Enter computer vision, the AI trying to out-count humans

It’s here that things get interesting for those of us thinking AI has figured it out already. In a 2020 paper titled “Automated Salmonid Counting in Sonar Data,” presented at the NeurIPS Workshop on Tackling Climate Change with Machine Learning, Kulits et al. of CalTech and Trout Unlimited developed a computer vision system aimed at automating much of this workflow by identifying and tracking salmon in sonar video with the goal of enabling round-the-clock monitoring across many rivers at once.

The catch: AI still misses roughly one in five fish

This is the humbling part of the story. The same 2020 paper reports that their fully automatic system achieved an error rate of 19.3% when compared against human counters. That’s a significant difference in an industry where a difference of thousands of fish one way or the other can make all the difference in changing the opening date of a fishing season.

Subsequent studies have reduced that number considerably. Kay et al. in a 2022 paper titled “The Caltech Fish Counting Dataset: A Benchmark for Multiple-Object Tracking and Counting,” presented at the European Conference on Computer Vision, found that newer object-tracking methods helped get the error rates down below 10%, and the authors claim automation can enable current sonar-based monitoring programs to scale from a few sites to entire watersheds.

Why this matters beyond the river

As noted in NOAA’s 2023 Arctic Report Card, Chinook and chum salmon stocks in Western Alaska have seen their lowest levels in decades amid a sharp rise in sockeye salmon stock levels. Scientists have yet to uncover the reason behind this phenomenon in light of climate change. This is why faster, cheaper, and more efficient monitoring cannot be regarded as simply an added benefit here; it can mean the difference between spotting a decline in population numbers early on or being surprised when things have gotten seriously out of hand.