Automatically detecting wallabies
150 years ago 3 breeding pairs of Bennett's wallabies were introduced to South Canterbury to start a fur trade. This has now grown reportedly to a population of almost 1 million animals and they're spreading, with sightings in the Mackenzie Country and North Otago common place.
The smaller Dama wallabies were introduced in the Rotorua area in 1912 and their population has spread throughout the Bay of Plenty and Waikato. There are other isolated populations of wallabies, such as on Kawau Island.
Wallabies are a pest, fouling pasture, damaging fences and crops. They contribute to erosion, kill plantation forest seedings and stop native bush regeneration. Earlier this year the government allocated $27m to wallaby control.
This year we have been working directly with the Otago Regional Council (ORC), with support from Environment Canterbury, to expand the capabilities of our thermal camera to automatically detect wallabies. Our thermal camera can already detect possums, rodents, cats, hedgehogs, mustelids, & rabbits. Wallabies seemed like a natural extension.
To enable this, our machine vision algorithms required examples of thermal videos of wallabies to learn from. Willowbank came to our aid - they graciously allowed us to set up our thermal camera in their wallaby reserve to capture footage and they repeatedly re-charged the battery over several months.
The next task was for someone to look at the videos we captured and tag those that contained wallabies. This task originally fell to ORC Environmental Officer Simon Stevenson. Luckily it coincided with some forced down time thanks to our first COVID lockdown. We now have over 2000 tagged videos of wallabies. Simon said he started dreaming in thermal videos. He was helped by Lincoln University postgrad student Saphy Hampshire.
With this set of training data the Cacophony machine learning engineer, GP was able to build a model that was capable of identifying wallabies with 94% accuracy. Check out their blog post with more technical details on how they built this wallaby recognition model. This video shows examples of the variety of wallaby footage we captured and of the model successfully identifying wallabies.
So we now have a thermal camera for sale that automatically detect wallabies. You can log into The Cacophony Portal, view the videos and find any that have had wallabies automatically identified. The next step in this project is to provide the option of receiving an email alert when a wallaby (or any other animal for that matter) is detected. This will allow you to respond to the detection in near real time, making it a tool that can help with control rather than just monitoring.
The first installation of this is on the Aviemore dam to see if we can detect wallabies crossing from South Canterbury to North Otago across the Waitaki river. It has only been a few weeks, but to date the only animal we have seen is (maybe) a stoat.
Thanks to all of those who contributed to the success of this project, including:
- Willowbank for the use of their wallaby reserve, including Kirsty's son who recharged the battery for us
- Simon and Saphy for their wallaby tagging
- GP for his work on the machine learning model
- ORC and ECAN for supporting the project
- Spark for supplying the mobile data and computing