Along Massachusetts’ rivers, an age-old event occurs every spring. Following some deep biological guidance toward freshwater spawning grounds, herring—small, silvery, and easily missed—begin to push upstream from the coast. For thousands of years, they have been doing this. However, their numbers have been declining recently, and for decades, those in charge of monitoring this decline have had to count fish with their eyes while standing at riverbanks.
The picture of a volunteer squinting at moving water while holding a clipboard reveals a lot about the state of conservation science. The work is necessary and honest. It’s also extremely constrained. It is only possible to count during the day. Volunteers grow weary. Additionally, the entire system silently collapses when hundreds of fish pass in a matter of minutes, a phenomenon known by researchers as a migration pulse. The figures soften. Data becomes untrustworthy. Furthermore, no one is truly aware of what occurs in the dark.
A group from the Woodwell Climate Research Center, MIT Sea Grant, MIT Lincoln Laboratory, MIT Computer Science and Artificial Intelligence Lab, and Intuit made the decision to try something new. Three Massachusetts rivers—the Coonamessett in Falmouth, the Ipswich River, and the Santuit in Mashpee—were equipped with underwater cameras, and a deep learning system that can recognize, track, and count river herring independently was developed. On the surface, the paper they published in February in Remote Sensing in Ecology and Conservation seems like a technical exercise. Beneath it, though, is something more significant.
The team was testing more than just a computer’s ability to count fish. They wanted to know if automation could close the particular gaps left by human observation. movement at night. brief pulses. the times when there are no volunteers and the river is free to do as it pleases. Maybe this is more important than anyone thought. What does the official count really mean if a fish population spikes between two and four in the morning and no one records it?
The findings were both genuinely fascinating and somewhat depressing. When used in different contexts, models trained on data from a single river in a single year performed poorly. More varied training data than the team had at first was needed for generalization, which is the capacity of a machine to apply what it learned in Falmouth to Ipswich. This finding suggests that no algorithm, no matter how well-designed, can easily overcome ecological variability, the messiness of real rivers, and real light conditions.
Nevertheless, the automated counts nearly matched human reviewers throughout entire migration seasons when the system was provided with sufficient diverse data. There was a significant efficiency gap. The computer processed the video continuously, without complaining, without pausing, and without missing a single overnight run—what would take an individual hours to review.
As this type of research advances, it’s difficult to ignore the issue of scale that citizen science has always attempted to address. There is only so much that a few skilled scientists can observe. Volunteers increase that reach, but their contributions are inconsistent. The arithmetic is completely altered by a machine that monitors a river day and night, flagging each fish that passes through the frame. The volunteers are not replaced by it. It covers the ground that they are unable to.
This spring, the herring are still in motion. They are most likely being recorded by a camera somewhere along the Santuit River. It’s still unclear if that data will ultimately influence a conservation strategy, management decision, or something that hasn’t been fully considered yet. However, the river doesn’t wait for assurance. It appears that the science doesn’t either.
