Thermal species classifier

Description

Wildlife Security Innovations and WildAP, the Wild Animal Phenotyping Infrastructure at the University of Helsinki, have been building technology together for wildlife research and coexistence. Our current project is a species classifier for thermal images. A camera that recognises the animal in front of it can act on its own, for example by playing a sound to a bear but not to a hare. That lets us run experiments on animal behaviour and communication in the field, day or night, without a researcher on site.

What we did

• Deployed WSI camera units to collect thermal training images
• Built a dataset covering 19 species and species groups
• Annotated every image by hand
• Trained the classifier on Roihu, Finland’s new supercomputer at CSC

Training the model

All the training images came from WSI’s camera units in the field, so the model learned from the same cameras it will run on. We annotated the images by hand, one by one, and trained the model on Roihu.

Results

With the classifier trained, we are now using it for research. We have fitted the WSI camera units with speakers and are putting them in the field for two studies. Bears and beehives.

In eastern Finland, the systems will detect bears that come close to commercial beehives. When a bear is detected, the unit plays a recording of people talking. We want to find out whether human voices are enough to keep bears away from the hives. Listening across species.

This winter, an MSc student will use the camera systems to study how mammals interpret the vocalisations of other species. Because the system knows which species is in front of it, it can play a chosen call to a specific listener and record the response.

To see some videos, check out our YouTube channel.

We have many great videos made with the system, such as this one.