How Machine Learning Can Aid Conservation of Migratory Birds

Juan P. Bello, Associate Professor of Music and Music Education, develops BirdVox-full-night, a new dataset of nocturnal migratory bird calls

Climate change, habitat loss, and human impact on the environment all pose threats to migratory birds. Consequently, avian conservation efforts depend on understanding the spatial and temporal distributions of bird populations. But since most bird populations migrate at night, avian researchers must typically rely on sound rather than sight. Human observation of nocturnal migratory birds, however, can be cumbersome and impractical.

To aid conservation efforts, CDS’ Juan P. Bello hopes that machine learning can automate the observation of migratory birds. While some automated methods already exist for recording and detecting nocturnal migratory bird calls, existing datasets are not well suited for machine learning research.

Bello, along with four other researchers, has developed BirdVox-full-night, a dataset including 62 hours of audio from 35,402 flight calls of 25 species of nocturnally migrating birds recorded by six sensors. The researchers placed ten ROBIN autonomous recording units around Ithaca, NY that gathered 966 recordings of which 548 are at least eight hours long.

With human annotations identifying flight calls from the full night recordings as a benchmark, Bello and collaborators trained four systems to detect bird flight calls: two energy-based detectors, a shallow learning pipeline, and a convolutional neural network.

To evaluate the accuracy of each system, the researchers balanced the dataset with 35,000 additional negative clips. They tested each method’s event detection capabilities on six full night recordings. The convolutional neural network outperformed the other methods with a 95% binary accuracy when augmented with additions of background noise, pitch transpositions, and time stretching.

The researchers point out that BirdVox-full-night, despite its state of the art capabilities, does have difficulty detecting rare flight calls due to its training. To mitigate this bias, BirdVox-full-night includes a test bed to allow for context-adaptive machine listening.

Bello and his team have released the BirdVox-full-night dataset through a Creative Commons license for fellow researchers and conservationists to access. As the first dataset of full night migratory bird recordings annotated by time and frequency, BirdVox-full-night will be a crucial tool for future avian conservationists.

By Paul Oliver

Like what you read? Give NYU Center for Data Science a round of applause.

From a quick cheer to a standing ovation, clap to show how much you enjoyed this story.