Machine learning models trained on forest sounds could help conservationists locate elusive possum survivors
New research shows that AI can scan thousands of hours of audio recordings to identify possum calls, aiding New Zealand's Predator Free 2050 program. However, the models often produce false positives when other animals make similar sounds. The study introduces a training method called 'cross-model confusion mapping' to reduce these errors.
The research applies machine learning to vast audio archives, automatically detecting the calls of possums—a key invasive species targeted by New Zealand’s Predator Free 2050 initiative. Because possums are nocturnal and elusive, acoustic monitoring offers a non-invasive way to track survivors across rugged terrain. However, the AI’s accuracy is undermined by false positives, as other native animals produce similar vocalizations. To address this, the study proposes “cross-model confusion mapping,” a training technique that teaches models to recognize and reject these overlapping sounds. This approach could refine automated wildlife surveys, making them more reliable for conservation planning without requiring constant human review.
This work could significantly improve how conservationists monitor endangered or invasive species, especially in remote areas where manual surveys are impractical. By reducing false positives, the method may lower costs and increase trust in AI-driven ecological data, potentially accelerating predator-eradication timelines. However, its impact depends on broader adoption and adaptation to other species and habitats. If successful, it could set a precedent for using machine learning in biodiversity management, though careful validation remains essential to avoid misdirecting limited resources.