Industries / Food, farming and forestry / Fishing and aquaculture
Finding and catching fish
Fishers find fish with sound and with floating rafts that fish gather under, then use gear changes, cameras and rules to keep unwanted animals out of the catch.
7
Milestones achieved, from 1935 to 2018.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1935
Echo sounders find fish
An echo sounder sends a pulse of sound down through the water and times the echo. K. Kimura in Japan published the first successful test of detecting fish this way in 1929, in small ponds; in 1935 Norwegian fisheries scientist Oscar Sund recorded spawning cod off Lofoten on an echo sounder, and after the Second World War European fishing boats began fitting echo sounders to find fish.
Credit
K. Kimura (1929 experiment); Oscar Sund (Lofoten, 1935)
1971
Dolphins let out of nets
In the eastern Pacific, yellowfin tuna often swim beneath dolphins, so purse seiners (boats that circle a school with a net and close its bottom like a drawstring purse) set their nets around dolphins. Captains developed a release manoeuvre called "backdown", believed first used in 1959, and in 1971 Captain Harold Medina fitted a finer-mesh panel in the release area so dolphins stopped tangling. Dolphin deaths fell from an estimated 550,000 in 1961 to about 1,000 a year since 2000, a cut of more than 99 percent, by the authors' count.
Lessona limit for each vessel worked because, in the authors' words, it placed "the fate of a vessel in its own hands".
Credit
Captain Anton Meizetich (backdown); Captains Harold Medina and Joseph Medina Jr. (panel); international dolphin death limits set for each vessel
1987
Turtle escape hatches required
A turtle excluder device (TED) is a metal grid sewn into a shrimp trawl; shrimp pass through the bars and turtles slide out through a flap. NOAA Fisheries began encouraging voluntary use in 1981, but few shrimpers took it up, so on 29 June 1987 NOAA made TEDs mandatory in the US shrimp fishery. NOAA reports that tests of top-opening designs cut turtle capture by 97 percent with little shrimp loss.
Lessonvoluntary uptake stayed low until a rule made the device routine.
Credit
NOAA Fisheries gear researchers working with shrimpers; the TED grew from a jellyfish excluder shrimpers already used
2006
Cameras audit the catch
Electronic monitoring (EM) means cameras and sensors on board that record fishing for review on land. In March 2006 British Columbia's hook-and-line and trap groundfish fleet, over 200 small boats, began 100 percent monitoring: cameras record every haul, and reviewers watch 10 percent of each trip's fishing and compare their counts with the skipper's logbook. A failed audit can mean full video review at the vessel's expense.
Lessonwatching all the video cost too much, so a random audit backed by a penalty did the job.
Credit
the BC groundfish industry, Fisheries and Oceans Canada and Archipelago Marine Research
2008
Buoys report fish below
A drifting fish aggregating device (FAD) is a raft that tuna gather under. Between 2008 and 2010 most eastern Pacific tuna skippers who fish on FADs adopted satellite buoys with echo sounders; each buoy reports an estimate of the fish beneath it, so boats visit only the full ones. By 2015, 55 percent of surveyed vessels had echo sounders on 75 to 100 percent of their buoys.
Lessonthe first buoys, around 1999 to 2000, gave poor estimates and skippers dropped them; uptake came with a better second generation in 2006 to 2007. _The-use-of-echo-sounder-buoys-in-purse-seine-fleets-fishing-with-DFADs-in-the-EPO.pdf).
Credit
survey by Gala Moreno, Jefferson Murua and Victor Restrepo (ISSF and AZTI); the survey does not name buoy makers
2014
Streamers keep albatrosses away
A bird-scaring line is a rope with bright streamers towed behind a boat so seabirds stay clear of cables and hooks. In South Africa's trawl fishery, which was killing some 10,000 seabirds a year, a seven-year study found the lines cut seabird deaths by 90 percent and albatross deaths by 99 percent since 2006; the lines are now mandatory there.
Credit
BirdLife South Africa's Albatross Task Force (leader Bronwyn Maree) and co-author Ross Wanless, published in Animal Conservation
2018
Lights steer fish from trawls
In 2014 Oregon scientists fixed LED lights to the footrope, the bottom edge of a shrimp trawl's mouth. Over 42 tows, bycatch of eulachon smelt (a threatened fish) fell 90 percent, juvenile rockfish 78 percent and flatfish 69 percent, with no significant loss of shrimp. Nearly all Oregon shrimp trawlers adopted the lights, and they became required in 2018.
Credit
Bob Hannah and colleagues, Oregon Department of Fish and Wildlife
Watch for
AI reads in-trawl video
Scientists adapted an object-detection model (YOLOv11) to find and label pollock and salmon in video from cameras inside trawl nets, footage used to judge salmon excluders (net panels that let salmon escape). The best model found 97 percent of fish with 82 percent prediction accuracy and processed a tow in hours, where people need days to weeks; it made many false salmon calls when herring were present.
Achieved whena group outside the team runs it on new tows and its counts match human reviewers, or managers use it for salmon bycatch decisions in the pollock fishery.
Claimed by
NOAA Fisheries Alaska Fisheries Science Center (study lead Katherine Wilson), September 2025
Cameras count catch on board
The Nature Conservancy (TNC) says an AI system on longline tuna boats reviews camera footage of the catch on board and sends daily summaries over Starlink; its development partner Tryolabs says it "achieves a 6% catch count error compared to expert reviewers." The prototype was tested in the eastern tropical Pacific, and Palau has backed a scale-up. In US camera programmes, people still watch between about 10 and 100 percent of the video, according to NOAA Fisheries in 2023.
Achieved whena fishery agency or independent reviewer publishes the count error on routine trips, or a regulator accepts the AI summaries for compliance checks.
Claimed by
The Nature Conservancy and Tryolabs, April 2026
Open
01
Check TED bars from photos
a phone workflow that photographs a TED grid next to a printed scale target, finds every gap between the bars, measures it, and flags any gap wider than the 4-inch (101.6 mm) legal maximum. It could also log the grid's angle with the phone's tilt sensor.
The briefCredit · unclaimed02
Flag unreviewable camera footage
a classifier that scans fishing-boat camera video before a human reviewer sees it and marks the stretches that cannot be reviewed, with a reason: lens covered by water or slime, glare, darkness, or crew blocking the view.
The briefCredit · unclaimed
Open 01 · the brief
Check TED bars from photos
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
a phone workflow that photographs a TED grid next to a printed scale target, finds every gap between the bars, measures it, and flags any gap wider than the 4-inch (101.6 mm) legal maximum. It could also log the grid's angle with the phone's tilt sensor.
Why it is still open
NOAA's gear team inspected 134 TEDs in 2024 (388 the year before) and named bent bars pushing gaps past 4 inches as the most common violation. Checks are done by hand with a ruler and an angle meter, and NOAA offers to train crews to check. A search found no photo-based checker.
Built with what exists
a smartphone camera; a printed ChArUco calibration board and the OpenCV library (Python) to turn pixels into millimetres and find bar edges; calipers for ground truth.
First step
ask a net shop or dock owner for permission to photograph and hand-measure five TED grids, and record each gap in a spreadsheet.
Done looks like
30 TED grids, every gap measured both by calipers and by the tool; the tool flags every gap the calipers put over 101.6 mm (zero misses), with a mean error of 3 mm or less; the measurement table and code published.
Open 02 · the brief
Flag unreviewable camera footage
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
a classifier that scans fishing-boat camera video before a human reviewer sees it and marks the stretches that cannot be reviewed, with a reason: lens covered by water or slime, glare, darkness, or crew blocking the view.
Why it is still open
in British Columbia's camera programme, reviewers found events they could not use because of "poor lighting" or "water/slime" on the lens, and at first the reasons were not recorded. A 2024 CSIRO study of longline camera video reports droplets and condensation on lenses causing blur and reflections. A search found no published tool or vendor feature with a measured miss rate.
Built with what exists
the Fishnet Open Images Database (86,029 images from cameras on tuna longliners); a waterproof action camera on a dock, with a spray bottle and lamps to make controlled fouling; labelling in CVAT or Label Studio; an off-the-shelf image classifier in PyTorch, with OpenCV's blur score as a simple baseline. Ready-made object detectors such as the YOLO family, which NOAA adapted for trawl video, are what make a small team's version practical.
First step
download Fishnet, hand-label 500 frames as reviewable or not with a reason, and publish the labelling rules.
Done looks like
a published, labelled test set of at least 2,000 frames from at least 20 cameras with reason codes; the classifier catches at least 90 percent of unreviewable frames while wrongly flagging no more than 10 percent of good ones; results reported for each reason.