Researchers develop one-shot learning system to detect calls from critically endangered pigeons

Automated bird-detection systems like BirdNET excel at identifying common species with thousands of training examples, but they stumble when it comes to endangered birds documented in only a few recordings. Researchers have now developed a workaround: a streamlined classification pipeline that works with minimal audio data and can monitor species on the brink of extinction.

The Challenge#

Conservationists tracking the last surviving members of critically endangered bird populations face a fundamental problem. Mainstream audio classifiers require substantial training datasets, which don't exist for species with just one, two, or three known calls on record. This leaves researchers without practical tools to detect and monitor these animals in the field—a serious limitation when every sighting matters.

How the System Works#

Rather than training a new classifier from scratch, the researchers leveraged the internal structure of existing large-scale bird classification networks. They built a lightweight detector using cosine similarity to compare new audio against a reference embedding space—a mathematical representation that captures the acoustic character of a species.

The pipeline incorporates preprocessing steps for filtering and denoising to optimize detection with sparse training data, plus targeted manual review checkpoints to maintain quality. This approach essentially allows the system to recognize a species by proximity in a learned acoustic space, rather than demanding hundreds of labeled examples.

Real-World Validation#

The team tested their method on the tooth-billed pigeon (Didunculus strigirostris), a critically endangered species from Samoa with no existing classifiers and only three confirmed recordings in scientific literature. Using this minimal dataset, the final model achieved 100% recall and 95% accuracy in identifying the pigeon's calls—strong enough for practical field deployment.

They validated the approach more broadly using a simulated scenario with diverse species from the Xeno-Canto audio archive, assessing different embedding spaces through clustering metrics to identify which performed best.

Implications for Conservation#

By releasing the system as open-source software, the researchers have created a template that conservation teams can adapt for monitoring other endangered species with scarce acoustic data. The technique bridges a critical gap between state-of-the-art deep learning and the real-world constraints of protecting species that barely leave a trace in the digital record.