Machine learning to classify animal species in camera trap images: Applications in ecology
Michael A. Tabak, Mohammad Sadegh Norouzzadeh, David W. Wolfson, Steven J. Sweeney, Kurt C. VerCauteren, Nathan P. Snow, Joseph M. Halseth, Paul A. Di Salvo et al.
Methods in Ecology and Evolution·2018·article·cited by 500
importance 6/10partially solvedWildlife Ecology and Conservationwildlife ecologystage: research
Problem
Camera trap studies produce millions of images that normally must be inspected manually to identify animal species.
Why it matters
Manual inspection of millions of images limits the scale of wildlife research and management.
What the AI did
Trained convolutional neural networks (ResNet-18) on 3,367,383 camera trap images from five US states to automatically classify wildlife species, tested on independent US images and out-of-sample Canadian ungulate images.
Results
not stated in source
Impact
Offers a scalable automated alternative to manual camera trap image analysis.
Record
Venue
Methods in Ecology and Evolution
Published
2018
Work type
article
Authors
Michael A. Tabak, Mohammad Sadegh Norouzzadeh, David W. Wolfson, Steven J. Sweeney, Kurt C. VerCauteren, Nathan P. Snow, Joseph M. Halseth, Paul A. Di Salvo et al.
camera traps, species classification, ResNet-18, wildlife ecology, CNN
Importance rating
6/ 10
significant — thousands to millions affected
Camera-trap studies produce millions of images needing manual inspection; automated classification (98% US accuracy, ~2,000 images/min) enables scalable wildlife research and management.
Basis: stated.
The source itself states the scale.
Stated scale: Millions of camera-trap images; 3,367,383 training images from five US states. Method and limits.