To test the assumptions underlying your remote camera analysis methods, you need animals. Lots of animals, mingling about in such a way that there are lots of detections in a short period of time. Hence, the ABMI found itself at the Edmonton zoo this past fall!
The ABMI uses a distance detection analysis to determine animal density from remote camera images; a crucial assumption of this method is that all animals who pass within 5 metres of the camera are detected. If this assumption is not met for a particular species, we would be underestimating its density. We did a pilot test of this assumption back in 2017 and wanted to revisit it to explicitly consider how body size related to detectability.
To test if every animal that passes within 5 metres of a motion-triggered camera triggers the camera, we paired time-lapse and motion-triggered cameras. The time-lapse cameras were set to take a picture every 3 seconds, which (hopefully) would provide us with a near-continuous record of the area being surveyed. We would then use these images to compare with what was caught by the motion-triggered cameras. But where to put the cameras? The time-lapse cameras typically either drained their batteries or filled up their memory cards within 48 hours, so this test needed to be located somewhere convenient for frequent resupply visits. In 2017, we opted for a semi-“wild” setting, and came up with only a handful of detections. Similarly, most ABMI core cameras will only record a handful of species detections over many months; it would be nearly impossible to maintain the time-lapse cameras over that duration of time.
Luckily, the Edmonton Zoo was happy to help us set up paired cameras in or adjacent to the enclosures of three different species: reindeer, arctic wolves, and red foxes. These species were chosen to cover a range of animal body sizes and behaviours, and were close matches to major mammal species found throughout Alberta. In October of 2020, we ran this experiment for ~20 days, servicing the time-lapse cameras every few days.
We ended up with ~410,000 images in total, of which almost 400,000 were time-lapse series. To weed out images without animals, we first used Microsoft’s Megadetector, a publicly available machine learning image classification model trained specifically on camera trap images. After sorting out why the model was 65-85% certain a shed in the wolf enclosure was, in fact, an animal, the winnowed images were uploaded to our tagging platform, WildTrax. The Megadetector model helped us cut down the number of images that needed manual tagging by ~90%!
Results and Next Steps
The consistency with which animals were detected by the motion-triggered cameras was related to body size: reindeer were more consistently detected than wolves, which were more consistently detected than foxes. Unsurprisingly, the speed at which animals moved through the detection zone also affected their detectability. With these results, we can be fairly confident that wild animals the size of reindeer (or larger) are well-detected within the 5m mark--moose, elk, woodland caribou, etc. Although white-tailed deer and mule deer are slightly smaller than our reindeer study subjects, their movement patterns are similar and we would expect that they are also consistently detected. However, this assumption may not hold as well for more intermediate-sized animals, such as wolves, lynx, or coyotes. To continue to refine our monitoring methods to better capture these smaller animals whose behaviour may influence their detection, we’re going to look for ways to improve detection: this includes testing out different camera configurations, such as placing a camera lower to the ground and even including a second motion-detection camera to survey the field of view. Stay tuned!
For more information, feel free to contact Marcus Becker at email@example.com
Project Team and Partners:
The Edmonton Valley Zoo