Point and Click

Posted on Tuesday, May 5th, 2020

Close-up of eye with computer code overlaid
With computer vision, a machine recognizes images in much the same way as humans. This technique can be applied to self-driving cars or facial recognition.

U of G researchers test the limits of artificial intelligence for analyzing ecological images.

Camera traps have captured incredible images of our planet’s elusive and endangered wildlife. Set up in the wild, these digital devices are triggered by an infrared motion sensor, snapping a photograph when animals move past. A single camera trap can accumulate thousands to millions of images. The challenge for ecologists is that someone must review and classify (label) the species for each image before ecological information, such as population dynamics, can be assessed. This process is laborious, expensive, and time consuming.

Artificial intelligence techniques, often referred to as deep learning, can be used to “train” computers to label images by learning features from images that have been labelled by humans. The computer can then apply what they learned to unseen examples. These models can be used to automatically analyze camera trap images, but there are challenges.

University of Guelph PhD candidate, Stefan Schneider, working in computer science professor Stefan C. Kremer’s lab, is examining the limitations and capabilities of deep learning systems. Collaborating with engineering professor Graham Taylor, a world-renowned expert in machine learning, and Prof. Saul Greenberg from the University of Calgary, Schneider examined a dataset provided by Parks Canada. He trained and tested a deep learning model considering 47,279 images collected from 36 unique geographic locations, representing 55 animal species. The goal was to help ecologists assess whether deep learning is an effective method for their dataset. The researchers compared six deep learning systems on the Parks Canada dataset considering “trained” and “untrained” locations. The team found that the machine learning systems performed well on species identification when backgrounds were seen during training and underperformed when they were faced with species in new locations. When considering locations seen during training, deep learning shows immense promise for modest-sized ecological datasets. The researchers have made their code publicly available.

“This study will inform ecological research efforts, which are crucial to understanding population dynamics of ecosystems across the planet,” says Schneider. “We recommend that ecologists have a minimum of 1,000 labeled images per species classification of interest as a training standard. This concrete recommendation will help ecologists determine when deep learning methods are appropriate.”

Deep Learning systems may get confused by new environments, like snow or new locations, and identify the bear incorrectly if not shown examples during training.



Graham Taylor holds a Tier 2 Canada Research Chair in Machine Learning.

Schneider S, Greenberg S, Taylor GW, Kremer SC. Three critical factors affecting automated image species recognition performance for camera traps. Ecol. and Evol. 2020 Mar 7. doi: 10.1002/ece3.6147

News Archive