Bird Data Gets a Big Boost

Recordings from northeastern national parks are building a foundation for the next generation of bird monitoring data analysis.

Every spring morning, forests across the northeastern United States fill with birdsong. Those sounds are more than part of the experience of being outdoors—they can tell scientists which bird species are present, when they are active, and how bird communities may be changing over time. Because bird populations respond to changes in habitat, climate, food resources, and other environmental conditions, long-term bird monitoring can reveal broader changes taking place in park ecosystems.

The Northeast Temperate Network, one of the National Park Service’s long-term ecological monitoring networks, has surveyed breeding birds in northeastern parks for many years. Park scientists and their partners monitor birds to better understand the health and changing character of park ecosystems. Increasingly, that work includes autonomous recording units, or ARUs: small devices that can record the forest soundscape for hours at a time. The recordings offer enormous potential for monitoring, but they also create a challenge: someone—or something—still has to determine which birds are making all those sounds.

Researchers recently used recordings from Acadia National Park, Marsh-Billings-Rockefeller National Historical Park, Katahdin Woods and Waters National Monument, and Hubbard Brook Experimental Forest to build a large, carefully annotated collection of bird sounds. The dataset can be used to develop and test automated tools for identifying birds by sound.

The collection contains 217 hours of annotated audio from 104 recording sites across four northeastern forest landscapes totaling more than 176,000 bird vocalizations representing 88 species.

Recording the dawn chorus

During the 2022 and 2023 breeding seasons, researchers deployed acoustic recorders in Acadia, Marsh-Billings-Rockefeller, Katahdin, nd Hubbard Brook Experimental Forest, a long-term ecological research site in New Hampshire. Together, the sites span a broad range of northeastern forest habitats and several hundred kilometers of geography.

The recorders collected more than 71,000 hours of audio—more than eight years of continuous sound—during the two field seasons. From that archive, researchers selected 1,302 ten-minute recordings for detailed annotation. Some were recorded at the same time that observers were conducting traditional point counts, in which an observer stands at a fixed location and records the birds they see and hear, while others came from carefully selected clear-weather mornings during the peak breeding season.

What does it mean to “annotate” a bird song?

For automated bird-identification tools to be trained and rigorously tested, researchers need recordings in which the sounds are reliably identified. That means showing exactly which species produced a sound and where and when it occurs in the recording.

Researchers opened each ten-minute recording as a spectrogram—a visual representation of sound—and listened while examining the patterns on the screen. When they confidently identified a bird, they drew a box around the vocalization and recorded its species, beginning and ending time, and frequency range. Overlapping birds were labeled separately.

The annotations went through three full stages of review. Later rounds were conducted by experienced local birders and ornithologists, checking identifications and adding sounds that earlier reviewers may have missed. By the final review, 98.6–99.9% of existing annotations were being retained, indicating that the dataset had become highly stable.

Creating Organization Out of a Chaotic Soundscape

Automated bird-recognition programs already exist, but training them to work reliably in real-world forest conditions is difficult. Many traditional sound libraries contain relatively clean recordings focused on individual birds. An ARU sitting in a forest hears something very different: a virtual cacophony of multiple birds singing at once, distant calls, mammals, insects, wind, and other background sounds.That difference matters. Models trained mainly on clean recordings may not perform as well when confronted with complicated real-world soundscapes. Expert-annotated ARU recordings provide something particularly valuable: a known ‘ground truth’ that can be used to train, benchmark, and test automated bird-identification models.

NPS image

Why this matters for parks

For parks, better automated bird-identification tools could eventually make it easier to turn thousands of hours of recordings into useful information about which birds are present and how bird communities are changing over time. Several recording locations at Acadia and Marsh-Billings-Rockefeller overlap with long-running National Park Service bird-monitoring sites, creating a direct connection between newer acoustic methods and observations made by people in the field.

The dataset is also publicly available to researchers developing and evaluating automated bird-identification tools. The new dataset is publicly available for researchers developing and evaluating automated species-identification tools. But its usefulness extends beyond teaching automated bird-identification systems, including machine-learning tools systems to recognize birds. The recordings and annotations can also support studies of bird occurrence, singing activity, geographic differences in song, interactions among vocalizing species, and the seasonal timing of bird song.

Related work is already underway using these recordings to investigate questions such as how acoustic detections compare with traditional point counts and how differences in bird songs affect automated detection. Together, these recordings and expert annotations could help scientists unlock far more information from the enormous—and rapidly growing—archives of bird sound being collected in national parks and other forests around the world.

Read the Study

Panwar et al. 2026, Scientific Data

Explore the Dataset

Audio recordings, annotations, and metadata are available on Zenodo.

Tags: acadia national park, katahdin woods and waters national monument, marsh - billings - rockefeller national historical park, netn, breeding landbirds, species spotlight, aru