Using machine learning to provide automatic image annotation for wildlife camera traps in the Arctic

Abstract

The arctic tundra is considered the terrestrial biome expected to be most impacted by climate change, with temperatures projected to increase as much as 10 °C by the turn of the century. The Climate-ecological Observatory for Arctic Tundra (COAT) project monitors the climate and ecosystems using several sensor types. We report on results from projects that automate image annotations from two of the camera traps used by COAT: an artificial tunnel under the snow for capturing information about small mammals, and an open-air camera trap using bait that captures information of a range of larger sized birds and mammals. These traps currently produce over two million pictures per year. <p>

<p>We have developed and trained several Convolutional Neural Network (CNN) models to automate annotation of images from these camera traps. Results show that we get a high accuracy: 97.84% for tunnel traps, and 94.1% for bait traps. This exceeds previous state of the art in animal identification on camera trap images, and is at a level where we can already relieve experts from manual annotation of images.