More than a third of all the world's bird species are harvested for sale in wildlife markets that include more individuals and species of birds than any other taxonomic group and are valued at tens of billions of dollars annually. Demand for live birds as pets and in cultural practices, such as bird singing contests and prayer animal release, is driving overexploitation, a significant contributor to population declines. In East and Southeast Asia, for example, high demand for captive songbirds is pushing some species to the brink of extinction in the wild.

Renowned for their beauty and songs, white-eyes (Zosteropidae) are popular songbirds in Asian wildlife markets and include endangered species whose trade is prohibited by CITES.

White-eyes sold in markets include endangered and near-threatened species, such as the Javan white-eye (Zosterops flavus) and Togian white-eye (Zosterops somadikartai), respectively. Protecting their wild populations is a critical part of preventing their extinction. Identifying such birds is crucial for enforcing wildlife laws and keeping them out of markets.

Researchers from the International Bird Conservation Partnership (IBCP) have found an AI-based solution to this problem. Many birds are highly vocal, communicating through various distinctive songs and calls. They realized that by harnessing recordings of their vocalizations, with their unique bioacoustic signatures, an AI tool could be developed to identify white-eye species from their vocalizations alone.

Leveraging public database

Public databases of bird sounds include xeno-canto and Cornell's Macaulay Library. They used three major avian vocalization databases to access bioacoustic data for 15 commonly traded, visually similar white-eye species. Using those recordings, they employed deep learning techniques and trained a powerful neural network model to recognize the specific acoustic patterns and sonic signatures of each white-eye species.

This approach is similar to how facial recognition is used to identify individual people, but they used vocal recognition to identify bird species instead. To make the system robust, they also incorporated data augmentation methods. They even included samples of ambient environmental sounds that might be picked up in market recordings in order to simulate real-world conditions as closely as possible.

Analyzing the results

The results of the study were published in Ibis. The ML models can identify focal species from their vocalizations with over 90% accuracy. That's an astounding precision for distinguishing between look-alikes based solely on sounds. However, the true power of this technology lies in its potential applications. For example, a user-friendly smartphone app using this system can be developed by anyone—law enforcement officers, customs agents, conservationists, and even citizen scientists. This would allow users to simply open the app, let it "listen" to the birds for a few seconds, and identify the species almost instantly, whether it's in a market, a pet shop, or in the field. While the study initially focused on white-eyes, it is adaptable for many vocal species.

Real-time data processing and action are invaluable for enforcing wildlife protection laws and clamping down on illegal trafficking. This bioacoustics approach offers an affordable, non-invasive, rapid, and highly accurate method for identifying birds solving the problem of distinguishing look-alike species.

Phys.org