Computer Vision System Automates Fish Population Counts During Spring Migration
Researchers at MIT and the Woodwell Climate Research Center have deployed underwater cameras and artificial intelligence to count river herring during their annual spring migration to Massachusetts spawning grounds. The system counted 42,510 fish in the Coonamessett River in 2024 and revealed migration patterns invisible to traditional monitoring methods.
River herring populations have declined severely over recent decades. Biologists rely on counting migrating fish to assess population health and guide conservation decisions. The problem: volunteer visual counts happen during daylight hours and miss nighttime movement, when hundreds of fish can pass within minutes.
The team built an automated pipeline using underwater video and deep learning models trained to detect and track individual fish across video frames. They labeled 1,435 video clips and 59,850 frames by hand, varying conditions to include different lighting, water clarity, fish species, and seasons.
The system's counts matched human video reviews, stream-side visual counts, and data from passive tracking tags. More importantly, it revealed behavior patterns: upstream migration peaked at dawn, while downstream migration occurred mostly at night-likely when fish avoid predators in darker conditions.
Scaling Beyond Manual Review
Traditional underwater video review is labor-intensive. Passive acoustic monitoring and imaging sonar work in some conditions but remain expensive. Manual video inspection, the cheapest option, requires hours of human time per season.
The computer vision approach processes continuous video feeds automatically, capturing migration timing and intensity that spot checks miss. Researchers trained models on multi-year, multi-site data, which performed better than single-location models.
Citizen Science Remains Essential
The researchers emphasize that volunteers still matter. Citizen scientists maintain cameras, label training data, and verify model outputs. The system supplements rather than replaces traditional counts.
Fisheries agencies haven't yet adopted automated counting at scale. Until they do, maintaining traditional monitoring ensures consistent long-term datasets for comparison. The two approaches together provide more complete population assessments.
The work was funded by MIT Sea Grant with additional support from the Northeast Climate Adaptation Science Center and the AI and Biodiversity Change Global Center.
For researchers working in environmental monitoring or applying machine learning to ecological data, this framework demonstrates practical computer vision deployment in field conditions. Learn more about AI for Science & Research applications in environmental work.
Your membership also unlocks:
