sammlapp/pa_forest_bird_localization_2023

Dataset_title

Creators: Sam Lapp (sam.lapp@colorado.edu)

Affiliations: University of Pittsburgh

Version number [v1]

DOI: tbd

General characteristics

audio format: 10 second clips beginning 4 seconds before variable-length localized events dimensions localized: 2 number of localization arrays: 3 array geometry: square grid with point centers sounds localized: songs of several Pennsylvania forest bird species number of audio files: 2784 size: 688 Mb

Metadata Coordinate Reference System (CRS): grid-relative coordinates in meters, relative to the Western-most and Southern-most limits of each microphone array;reference coordinates of 0 m East and 0 m North for each array are given in ./localization_metadata/grid_origins.csv in the CRS EPSG:32617 (UTM Zone 17S).

Audio clips were extracted starting 4 seconds before each localized event. The clips from a single localized event are extracted at the same real-world timestamp. The TDOA estimate is incorporated into the start_time field: this field will be 4.0 seconds for the microphone with the earliest detection, and larger than 4.0 by the tdoa for other microphones. The duration field corresponds to a manual reviewer's annotation of the section of the song that was not obscured by other songs (clips in which the song was masked were excluded from the dataset).

Abstract

This dataset contains audio and metadata for acoustically localized bird songs from Pennsylvania forests. We used acoustic localization grids to acoustically localize the positions of forest birds and capture redundant recordings of each song at varying distances. The data was used to investigate singing volume and habitat attenuation in the context of acoustic distance estimation and quantifying the active space of songbird vocalizations in passive acoustic monitoring datasets. A manuscript describing these analyses is in review.

This dataset provides 10-second audio clips of acoustically localized songs for 12 bird species, with 2-84 events per species. Automated procedures were used to perform species detection and acoustic localization. Acoustically localized events were filtered based on residual error and spatial position estimate. A manual review process confirmed the species identity and checked for for correct temporal alignment, a requirement for correct acoustic localization. Each song is linked to multiple audio recordings, taken by the microphones which were used to estimate the spatial position of the event. Audio clips where the target sound was masked were removed from the dataset. The audio clips are accompanied by a table describing the positions, species, and timestamps of each localized song.

Study description

This study used acoustic localization grids in Southwestern and Central Pennsylvania to generate acoustically localized positions of forest birds with redundant measurements of each song at varying distances. The data was used to investigate singing volume and habitat attenuation in the context of acoustic distance estimation and quantifying the active space of songbird vocalizations in passive acoustic monitoring datasets. A manuscript describing these analyses is in review.

Personnel: Lapp, Sam (sam.lapp@colorado.edu). University of Pittsburgh. https://orcid.org/0000-0003-1637-6822 Rhinehart, Tessa. University of Pittsburgh. https://orcid.org/0000-0002-4352-3464 Knight, Elly. Boreal Avian Modelling Centre; Department of Biological Sciences at the University of Alberta. https://orcid.org/0000-0002-8578-892X Edwards, Brandon. Institute for Wetland and Waterfowl Research. https://orcid.org/0000-0003-0865-3076 Yip, Dan. Environment and Climate Change Canada. https://orcid.org/0000-0002-7104-6482 Bayne, Erin. Boreal Avian Modelling Centre; Department of Biological Sciences at the University of Alberta. https://orcid.org/0000-0002-0679-4521 Kitzes, Justin. University of Pittsburgh. https://orcid.org/0000-0001-7839-3594

Data types collected: synchronized acoustic recordings, acoustically localized positions of bird songs, species ID of localized songs

Species List

The following species were detected using the BirdNET classifier and localized using OpenSoundscape. We only retained songs during the review process, discarding localizations of calls or other sounds. The English common name is used for the label column in localized_events.csv. Species names follow the Avilist 2025 taxonomy. These data are also available in the classes.csv file.

Common Name	Localized Song Count	Scientific Name
Blue-headed Vireo	84	Vireo solitarius
Black-throated Green Warbler	81	Setophaga virens
Hooded Warbler	72	Setophaga citrina
Scarlet Tanager	64	Piranga olivacea
Wood Thrush	62	Hylocichla mustelina
Black-and-white Warbler	58	Mniotilta varia
Chestnut-sided Warbler	55	Setophaga pensylvanica
Eastern Wood-Pewee	55	Contopus virens
Ovenbird	49	Seiurus aurocapilla
Eastern Towhee	47	Pipilo erythrophthalmus
Hermit Thrush	3	Catharus guttatus
Eastern Whip-poor-will	2	Antrostomus vociferus

Files

localized_events.csv: table with a row describing each acoustically localized sound (the table is also provided as localized_events.parquet to facilitate proper loading of timestamps and columns containing lists).

Columns: event_id: unique identifier for each acoustically localized sound within the dataset label: species localized (see classes.csv and description above) start_timestamp: onset time of the event in ISO format (e.g., 2025-05-20T10:00:00.000-00:00). Note that start timestamps are given in UTC, while the local timezone is Eastern Daylight Time (UTC-5). duration: event length in seconds x,y: estimated position of the sound source in meters; given relative to the Western-most microphone coordinate and Southern-most microphone coordinate; note that 'z' is empty because localization was performed in 2 dimensions file_ids: list of file_id for audio clips that participated in the localization - maps to file_id column of audio_file_table.csv file_start_time_offsets: time in seconds from start of each audio file listed in file_idsto the start of the event

classes.csv: table of species names and number of events per species. Columns: class: English common name, Avilist 2025 taxonomy; these values are used in label column of localized_events.csv event_count: number of localized events in the dataset scientific_name: Avilist 2025 scientific name low_f: lower frequency of bandpass filter used prior to cross correlation during acoustic localization, in Hz high_f: upper frequency of bandpass filter used prior to cross correlation during acoustic localization, in Hz

/localization_metadata/audio_file_table.csv: lists all audio files with a mapping to the microphone metadata (point_table.csv) using the point_id column. Columns:

• file_id a unique identifier for each audio file; maps to file_ids in localized_events.csv
• relative_path: path from the top level directory of this dataset to the audio clip; for example: audio/pa_2023_loca_00008.FLAC
• point_id identifies the microphone that recorded the audio clip; maps to point_id in point_table.csv
• start_timestamp: the start time of the audio clip in ISO format (e.g., 2025-05-20T10:00:00.000-00:00). Note that start timestamps are given in UTC, while the local timezone is Eastern Daylight Time (UTC-5).

/localization_metadata/point_table.csv: lists the positions of each microphone; point_id column maps to point_id in audio_file_table.csv. The coordinates (columns x, y) are given in meters, relative to the Western-most microphone coordinate and Southern-most microphone coordinate. The coordinates are also provided in the CRS EPSG:32617 (UTM Zone 17S) in the utm_E and utm_N columns and the elevation is given in meters in the ortho_height column. The high_or_low columns is low except for microphones that were deployed directly above another microphone, i.e., on the same tree, in which case the value is high.

clip_table.csv: This table presents the localization metadata in alternative format, with one row per audio clip (rather than one row per localized event). Note that multiple rows correspond to a single localized event, as multiple microphones are used to perform a single acoustic localization. Columns: 'event_id', 'array_id', 'x', 'y', 'distance_m', 'label', 'start_time', 'point_id', 'duration', 'start_timestamp', 'file_id', 'rel_path', 'z'

event_id: unique identifier for each localized event (matches event_id in localized_events.csv) array_id: unique identifier for each microphone array (matches array_id in point_table.csv) x: estimated Easting position of the event in meters (grid-relative CRS) y: estimated Northing position of the event in meters (grid-relative CRS) distance_m: the distance from the microphone to the estimated sound source position, in meters label: the common name of the species, matching the label column in localized_events.csv and described in classes.csv start_time: the start time of the event (4.0 + tdoa) in seconds relative to the start of the audio file point_id: the unique identifier for each microphone (matches point_id in point_table.csv) duration: the duration of the annotated, unmasked singing event in seconds start_timestamp: the start time of the clip in ISO format file_id: the audio file associated with the event rel_path: the path of the audio clip, relative to the dataset root directory z: empty column (height of the sound source was not estimated)

/audio/: Contains 10-second FLAC audio clips from each microphone for each localized event, with all clips from an event extracted starting at the same timestamp (4 seconds before earliest event start time on any microphone in the event).

Scripts

Scripts in this repository are included for reference, rather than to reproduce the exact localization parameters in the included tables. Running the scripts requires access to the large, unpublished passive acoustic monitoring dataset. The exception is that the review.ipynb notebook can be run on the included data from this repository to visualize temporal alignment of localized songs.

/scripts/sync_utils.py: python script with functions to synchronize audio files

/scripts/1*sync_loca*.py: python scripts to accurately time-synchronize AudioMoth audio files by interpolating and resampling with GPS-recorded timestamps (runs on the raw, original PAM dataset, not included in this repository).

/scripts/2*_predict_*.py: apply BirdNet v2.4 species classification CNN to audio files to detect songs of target species across the full, original PAM dataset (not included in this repository).

/scripts/3*aggregate_dets*.ipynb: select detections from the automated species classifier to user for acoustic localization. Applies classifier score thresholding and selects commonly-detected classes for localization.

/scripts/4*localize*.py: python notebook using OpenSoundscape to perform acoustic localization on selectred species detections.

/scripts/loca_utilities.py: utilities for localization process

scripts/5_review.ipynb: notebook to visualize and review acoustically localized events; review process involves confirming that the detected sound was a song of the labeled species, and that cross correlation correctly captured the relative time of arrival of the song across microphones.

/scripts/6_create_public_dataset.ipynb: notebook used to generate this dataset from the full dataset and localization results

Sites

State Game Lands 111 and State Game Lands 34, Pennsylvania, United States

Hardware

Recorder source: Open Acoustic Devices Recorder model: AudioMoth 1.1.0 with GPS receiver Firmware version: AudioMoth GPS Firmware 1.0.8

Recording properties

Date range of data: 2023-05-17 to 2024-06-03

Recording schedule description: Times of recording: 06:00 to 11:00 local time (10:00 to 15:00 UTC) Sleep-wake schedule: continuous recording Sample rate: 24 kHz (resampled from original 48 kHz recordings) Other relevant settings: medium gain, 16-bit resolution, FLAC format (originally .WAV)

Data aggregation notes: After acoustic localization, we randomly selected localized singing events for manual verification, then performed manual review.

Recorder positioning

Placement

Spatial pattern or geometry: 50 m square grid with additional recorders at the center of each square

Dimensions of array: approximately 200 x 200 m (grid of 4x4 50m squares + centers)

Deployment: Recorders were placed in plastic zip-lock bags for environmental protection and zip tied to small trees (diameter at breat height < 10 cm when possible).

Position measurement

We measured microphone positions with a pair of SparkFun RTK GPS units in base-rover mode, which achieved 20 mm estimated accuracy. However, we observed that the measured positions would drift up to 5 meters over time.

Synchronization

Synchronization type: GPS onboard Audiomoth

Synchronization frequency: GPS timestamps taken every 1 second

Synchronization method: resampling for interpolation to the nominal sampling rate, using GPS timestamps recorded every 1 second.

Recording start/end time trimming: Synchronized recordings taken simultaneously were not trimmed to the same start/end time, but instead kept at their original durations.

Scripts: scripts/1*_sync*.py and scripts/sync_utils.py

Sound detection:

We used the BirdNET species classifier to detect songs of forest songbirds in Pennsylvania. Audio was not denoised or otherwise pre-processed. We detected and localized songs from several bird species that breed in Pennsylvania forest habitats (listed in classes.csv).

Detection strategy: BirdNET species classification CNN

Detector name and version: BirdNET v2.4 [1]

Link to detector information: https://github.com/birdnet-team/BirdNET-Analyzer

Post-processing detector outputs: We used a universal score threshold of -1 (logit score) for species detections, and excluded several non-local species from the BirdNET detections. We then examined species detection rates to select the species included in this dataset. We did not manually verify species detections before localization, as we later perform manual review of localized events.

Scripts/resources: scripts/2*_predict_*.py and scripts/3*_aggregate_dets_*.ipynb

Localization:

Time-delay-of-arrival based acoustic localization was performed in OpenSoundscape v0.11.0.

Localization algorithm:

• tdoa optimization with SoundFinder

Time delay calculation:

• Cross correlation algorithm: GCC-PHAT
• Filter: bandpass filter is applied prior to cross correlation, using species specific frequency ranges listed in classes.csv

Error rejection parameters:

• Minimum number of microphones with detection to attempt localization: 5
• Maximum distance from one receiver to adjacent receivers when grouping detections: 40
• Cross correlation threshold: 0.02 (if the maximum cross correlation value is below this threshold, the microphone's audio is not used in the event's localization)

Post-localization filtering:

• Positions with the root-mean-square residual of time delay (in meters) greater than 1 meter were removed.
• Positions whose height estimates were below -3 m or above 30 m were removed.

Scripts: scripts/4*_localize_grid*.py

Manual review:

We manually reviewed temporally aligned audio segments from each acoustic localization event to confirm that (1) cross correlation aligned the sound of interest rather than another sound; (2) the detected sound was a song of the labeled species (calls were rejected); and (3) to select the temporal portion of the localized audio clip at which the focal species song was louder than any other sounds in the soundscape. Audio files were removed from the localized event if other sounds were of equal or louder volume compared to the target sound, within the frequency range of the target sound. The selected time window is the window for which remaining clips had the target species song as the loudest sound in the species' frequency range.

Scripts/resources: scripts/5_review.ipynb

Observational data

None included in this dataset. Contact the authors for observed bird positions from the same localization grid; note that the singing events from the observed birds were not necessarily the same events that were randomly selected for review and inclusion in this dataset.

Acknowledgements

We thank Chapin Czarnecki for assistance with field data collection. This work was supported financially by NSF and the Gorgon and Betty Moore Foundation. We thank the Pittsburgh Supercomputing Center and the Center for Research Computing at the University of Pittsburgh for resources provided.

License

Creative Commons license - Attribution - CC BY 4.0.

References:

[1] Kahl, Stefan, et al. "BirdNET: A deep learning solution for avian diversity monitoring." Ecological Informatics 61 (2021): 101236.