A million photographs, one patient algorithm: AI learns to spot an endangered toad in the dark

Every winter, Western Leopard Toads cross a Cape Town road through a set of tunnels built for them — and camera traps photograph everything that moves. The result was an archive no human could read: more than a million photographs. Master's researcher Joshua Salako, at the African Institute for Mathematical Sciences in South Africa, working with Nature Connect, trained an AI model from just 20 labelled photos and set it loose on the archive. It found 730 toad sightings across 387 tunnel events from 2023 to 2025, including 15 repeat crossings by recognisable individuals. One toad, T2026-7, was matched crossing tunnel 4 in eight minutes on an August night, its identity confirmed on 35 of 36 back-pattern points. The open-source tool, AmphiLens, is now published for anyone to use — and this season alone it has already checked over a million more photos. An endangered species, a road, and a million pictures nobody had time to look at: now somebody — something — is looking.

SDG 15 · Life on land — Mechanism: tech innovation