AI Night Cameras Now Spot Deer and Boars Eating Crops

In Japan's satoyama — places where farms and forest sit side by side — deer, wild boar and monkeys cause serious crop damage every year. Farmers can't patrol all night, so "camera traps" (motion-triggered cameras that photograph whatever walks past) do the watching. The problem is that someone still has to sort through thousands of dark, blurry night photos to figure out what showed up and where.

A team of Japanese researchers released SatoyamaCT, a public dataset of 12,642 nighttime infrared photos from three rural regions, each one labeled by wildlife experts. Infrared means the cameras shoot in the dark, using light people can't see. The labels were built with an AI assistant that pre-sorted the images, cutting expert workload while keeping human oversight. Expert agreement on species labels was very high, around 90%.

The interesting part is honesty about failure. The AI performs well in some regions but drops 16 to 27 percentage points in Wakayama, one of the three areas — a gap that persisted across four different AI models and several standard fixes. So the team added a safety net: selective prediction, where the AI flags its own uncertain photos for human review. At 50% coverage — meaning the AI handles half the images on its own — Wakayama accuracy climbed from 59% to 82%.

Why care outside Japan? The same recipe applies anywhere wildlife, livestock or pests need round-the-clock monitoring: cheap cameras plus AI triage plus a human checking the doubtful cases. It's a realistic model for how AI actually gets used in the field — not perfect, but good enough to save enormous amounts of human time.