With funding from Predator Free 2050 and the Ministry of Business, Innovation and Employment (MBIE)’s Kiwi Rescue programme, Manaaki Whenua researchers Dr Brent Martin and Dr Al Glen have developed a free online tool, CamTrapNZ, which allows users to manage their images quickly and easily. The tool uses artificial intelligence (AI) to identify images of 15 taxa commonly detected by camera traps in Aotearoa New Zealand: kiwi, other birds, cats, deer, dogs, ferrets, goats, stoats/weasels, possums, rodents, hedgehogs, rabbits/hares, wallabies, pigs, and livestock. The model could be trained to recognise more taxa in the future.
The accuracy of species identification depends on the species, and also whether the image is colour (daytime) or black and white (night), says Al. “For example, the AI will detect wallabies around 98 percent of the time, but has much lower accuracy for mice. This is mainly because mice are so small that they often appear as little more than a dot in the images. Accuracy is generally higher with colour images than with black and white ones.”
Some species are grouped together because the images captured by camera traps don’t always show distinguishing features. For example, stoats and weasels can easily be distinguished if the tail is visible (e.g. images a and b below), but sometimes the images don’t show the tail (e.g. image c). Rats and mice are particularly difficult to separate.
Results can be viewed in a web browser or in a spreadsheet, which shows the species identified in each image, as well as a confidence score. This allows the user to sort images according to the level of confidence. For example, users might choose to manually check any images with a confidence rating below a certain level. In addition, the online tool allows users to produce maps, graphs, and other reports of their results. The software is hosted on TrapNZ, a free online platform widely used for recording pest control, monitoring, and biodiversity outcome data in Aotearoa New Zealand.
Further improvements to CamTrapNZ are planned, subject to funding availability. For example, Brent and Al hope to improve the accuracy of species identification by training the AI with larger numbers of sample images. “We also hope to add more species, and to add the capability for the system to improve over time by constantly retraining from the images users upload. We are also building a version of the software that runs on the user’s own computer, without the need to upload images to a server. This will be useful for users with very large numbers of images,” says Brent.
