Teaching AI to Ignore the Birds: Finding New Zealand's Last Possums by Sound

Removing most possums from a patch of New Zealand bush is the straightforward part. Finding the last few is where eradication gets hard, and where a handful of survivors can quietly rebuild a population. Writing in The Conversation, Akbar Ghobakhlou, a senior lecturer in the Department of Data Science and AI at Auckland University of Technology, describes a way through that bottleneck: overnight microphones paired with AI that scans thousands of hours of audio for possum calls. His team’s new research shows the approach can work well, “but only if the AI learns not to mistake other animals for possums.”

That caveat is the whole story. The AI models compact enough to run on small, battery-powered recorders in remote forests, without uploading huge audio files, are also the ones most prone to false alarms, wrongly attributing another animal’s call to a possum. For a conservation team, Ghobakhlou writes, “a false alarm can mean travelling to remote locations in search of an animal that isn’t there.” At the mop-up stage, when only a few animals remain, that wasted effort is the difference between a successful eradication and a stalled one.

The fix his team developed has an unusual shape. Rather than asking the AI directly which sounds were possums, they first ran the audio through BirdNET, a widely used system trained to identify more than 6,000 bird species but never trained on possums at all. That turned out to be the point. Because BirdNET could only classify a sound as some bird, “every possum call was forced into the bird species it most closely resembled,” revealing which birds an AI was most likely to confuse with a possum. To human ears the calls sound quite different, but the models work from spectrograms, visual representations of sound frequencies over time, where the patterns are more alike than they seem.

The team then fed recordings of those confusable birds back into training as “hard negatives,” examples the AI found difficult to tell apart from possums but needed to learn were not. Ghobakhlou likens it to teaching someone to distinguish two similar-looking people by repeatedly showing them both. They called the method “cross-model confusion mapping.”

The significance shows up in testing. Models trained this way “produced far fewer false alarms while maintaining high detection accuracy,” while comparable models trained without the technique “generated hundreds of false detections when analysing forest recordings that contained no possums at all.” One detail matters for anyone reading AI claims: that gap only appeared when the models were tested on recordings they had never encountered before. On familiar-style recordings, both approaches looked about the same. The method also held up on forest audio containing bird species the model had never heard in training, which suggests it learned to recognise possum calls more generally rather than memorising a few confusing species.

Ghobakhlou is careful about the limits. The native ruru, or morepork, was not flagged by BirdNET as a likely source of confusion, and how the system behaves around that nocturnal owl is left to future work. Further field testing is needed before wide deployment. But the same approach could potentially be adapted to stoats, rats and other invasive species targeted under Predator Free 2050, the national programme aiming to remove the worst introduced predators. As Ghobakhlou puts it, if AI can help teams “spend less time chasing false alarms and more time locating real pests,” it becomes another tool in that effort.

The full piece, including the prototype listening device and the underlying study, is worth reading in The Conversation.