Peninsula Road cuts straight through Cape Town’s False Bay Nature Reserve, right across a breeding ground of the endangered Western Leopard Toad, a species found only in the Western Cape. Every rainy winter the toads cross to breed, and cars kill them. So the nonprofit Nature Connect built six tunnels under the road, two in 2021 and four in 2023, and put cameras at the ends. Toads are cold-blooded and don’t trip the heat sensors that usually trigger camera traps, so these cameras take a photo every 15 seconds, all night. That adds up to a mountain of pictures of empty tunnel floor.
The clever bit
Joshua Salako, a master’s student at the African Institute for Mathematical Sciences (AIMS) South Africa, trained a toad detector, AmphiLens, that starts from just 20 hand-labelled photos. It sharpens the contrast of dark night frames, tries four detection models, and picks which photos people should label next so it learns fastest. Llewellyn Williams of Nature Connect built it into the team’s staff portal, where a person confirms every photo it flags. A second tool compares each toad’s lifelong pattern of dark, yellow-edged patches to find the same animal at both ends of a tunnel.
Those 384 photos add up to 130 toad visits once repeat shots taken minutes apart are counted once. The busiest night, 17 August, brought 16. Since October the system has checked 2,160,620 photos from the 2023 to 2026 seasons. On 21 August a toad the team calls T2026-7 was photographed at one end of tunnel 4 at 00:38 and at the other end eight minutes later, and 35 of the 36 points the computer matched on its back line up.
Hooray
A small conservation team can now show its tunnels at work, one toad at a time, without anyone reading a million photos by eye. Nature Connect says just one toad has died on the tunnelled stretch of road in three years.
What comes next
The team is still confirming this season’s matches and plans to share how many separate toads made the crossing once the checks are done. Salako has published AmphiLens openly, with its code, trained models and an install guide, so any camera-trap project can use it.