Biodiversity Data Journal :
Data Paper (Biosciences)
|
Corresponding author: Matthew James Weldy (matthewjweldy@gmail.com)
Academic editor: Therese Catanach
Received: 05 Jan 2024 | Accepted: 15 Apr 2024 | Published: 29 Apr 2024
This is an open access article distributed under the terms of the CC0 Public Domain Dedication.
Citation:
Weldy M, Denton T, Fleishman AB, Tolchin J, McKown M, Spaan RS, Ruff ZJ, Jenkins JMA, Betts MG, Lesmeister DB (2024) Audio tagging of avian dawn chorus recordings in California, Oregon and Washington. Biodiversity Data Journal 12: e118315. https://doi.org/10.3897/BDJ.12.e118315
|
Declines in biodiversity and ecosystem health due to climate change are raising urgent concerns. In response, large-scale multispecies monitoring programmes are being implemented that increasingly adopt sensor-based approaches such as acoustic recording. These approaches rely heavily on ecological data science. However, developing reliable algorithms for processing sensor-based data relies heavily on labelled datasets of sufficient quality and quantity. We present a dataset of 1,575 dawn chorus soundscape recordings, 141 being fully annotated (n = 32,994 annotations) with avian, mammalian and amphibian vocalisations. The remaining recordings were included to facilitate novel research applications. These recordings are paired with 48 site-level climatic, forest structure and topographic covariates. This dataset provides a valuable resource to researchers developing acoustic classification algorithms or studying biodiversity and wildlife behaviour and its relationship to environmental gradients. The dawn chorus recordings were collected as part of a long-term Northern Spotted Owl monitoring program; this demonstrates the complementary value of harnessing existing monitoring efforts to strengthen biodiversity sampling.
This dataset of dawn chorus soundscape recordings is one of the few open-access acoustic datasets annotated with non-biotic and both interspecific (across species) and intraspecific (within species) bird, mammal and amphibian sonotypes and the first that is paired with climatic, forest structure and topographical covariates extracted at recorder locations. This makes it a valuable resource for researchers studying the dawn chorus and its relationship to the environment.
annotated soundscapes, forest ecology, amphibian, mammal, bird, vocalisation
Scientists worldwide are documenting unprecedented and potentially accelerating declines in ecosystem health and biodiversity (
To overcome these limitations, monitoring and research programmes increasingly adopt innovative protocols that leverage high-throughput sensor technologies like autonomous recording units and motion-activated cameras (
The quality of labelled datasets is essential for the successful training, evaluation and generalisation of perception models (
The Northwest Forest Plan (hereafter NWFP;
In this context, we present annotated passive acoustic monitoring data collected during 2022 to support long-term monitoring of federally threatened spotted owl populations under the NWFP Effective Monitoring Program. The data were collected with Wildlife Acoustics Song Meter SM4 autonomous recording units during the hour following sunrise at 525 sites in California, Oregon and Washington, USA (Wildlife Acoustics, Maynard, MA). Additionally, we obscured exact sampling locations to protect sensitive species, but we provided 48 forest-structure-related environmental covariates extracted at each recorder location. This dataset provides value for researchers involved in developing or evaluating acoustic classification algorithms and for those interested in exploring the spatial variation in species-specific vocalisation phonology or the relationships amongst occurrence, vocalization behaviour and environmental characteristics, and contributes to making ecological research more transparent and reproducible (
We focused on the avian dawn chorus, including both migratory and resident species. The dawn chorus, characterized by the singing of numerous birds during the early morning hours, is an ecologically important period for studying avian behaviour (
In 2022, following the protocol outlined in
Four Song Meter 4 (SM4) autonomous recording units were deployed in a standardised arrangement in each hexagon. The recorders were positioned at least 500 m apart at a minimum distance of 200 m from the hexagon boundary. The SM4 devices were mounted on to small trees (15–20 cm in diameter at breast height) approximately 1.5 m above the ground on mid-to-upper slopes and ≥ 50 m from roads, trails and streams. The SM4 devices have two built-in omnidirectional microphones with a signal-to-noise ratio of 80 dB, typical at 1 kHz and a recording bandwidth of 0.02–48 kHz. The recording rate was set to 32 kHz at 16-bit resolution and the data were saved in uncompressed WAV format.
Recordings were collected for six weeks, from March to August. Each device recorded approximately 11 hours of audio daily. The daily recording schedule comprised a 4-hour window starting two hours before sunrise and ending two hours after sunrise, another 4-hour window starting one hour before sunset and ending three hours after sunset and 10-minute recordings at the start of every hour outside the two longer recording blocks.
Dawn chorus dataset preparation
We limited the available set of recordings to only those occurring from May–August to ensure we included migratory species (
Annotation methods
Prior to labelling, we developed a sonotype library, which describes the acoustic properties of 260 sound types and describes a standardised label structure (see dataset: metadata.tsv). Sonotype descriptions and categories were developed by acoustic and visual inspection of examples and supplemented by descriptions provided by
Two trained annotators labelled each 2-second window with labels from the set of potential sound types. Unknown signals were labelled ‘unknown,’ and clips with no biotic signals (or noise classes of interest documented in metadata.tsv) were labelled ‘empty.’ Windows were labelled ‘complete’ and considered fully annotated when every signal was assigned an annotation. Files were deemed fully annotated when every 2-second window of the 5-minute recording was assigned the ‘complete’ label. The label 'impossible' was utilised for instances where a biological sound was present in the window, but could not be confidently identified, often due to faintness or being obscured by rain. Additionally, eight aggregated biotic sounds were not separated due to uncertainty in assigning a label confidently.
We used two annotation methods: linear annotation and model-assisted annotation. The linear annotation method consisted of fully annotating dawn chorus recordings in sequence. The model-assisted method used BirdNET version 2.2 (
Environmental characteristics
We included 48 variables, including three climatic variables, 38 forest structure variables, five topographic variables and two masked spatial variables. We downloaded estimates of annual precipitation (mm), minimum temperature (°C) and maximum temperature (°C) averaged from 1970–2000 at a 1 km2 resolution from WorldClim version 2.1 (
Analytical Methods
We performed two sets of analyses. The first set of analyses used t-distributed stochastic neighbour (
We used acoustic recordings collected from federally managed lands in California, Oregon and Washington, focusing on forest-capable areas (Fig.
We took measures to protect sensitive species that might occur at our sampling locations. As a result, we obscured specific recording locations to the resolution of the overlapping Townships and Ranges, which are approximately 1 mi2 (2.58 km2) grid cells used by the U.S. Public Land Survey System. We also provide detailed environmental characteristics extracted from the actual recorder locations. This approach ensures data privacy, while allowing us to furnish essential information for our study.
37°43'48''N and 49°01'48''N Latitude; 125°00'00''W and 120°30' 00''W Longitude.
We identified 116 sound types during the annotation of these recordings (Table
This dataset includes annotations for 116 sound categories, including 58 avian species, two mammalian species, one amphibian species, eight aggregated biophonic sounds, one geophonic sound type and six anthrophonic sound types. Each annotation is accompanied by its corresponding sound type, common name, scientific name and species code, following the 2021 eBird conventions for Clement's taxonomy. Astericks (*) indicate novel class labels following Clement’s naming conventions.
Type |
Common name |
Scientific name |
eBird code |
Aggregated Biophonic |
Chipmunk |
Neotamias spp. |
chipmu* |
Drum |
Picidae |
drum* |
|
Fly |
Insecta |
fly* |
|
Parulidae complex |
Parulidae |
paruli* |
|
Setophaga complex |
Setophaga spp. |
setoph* |
|
Tree creak |
tree* |
||
Unknown chip |
Aves |
unk* |
|
Wingbeat |
Aves |
wingbeat* |
|
Amphibian |
American Bullfrog |
Rana catesbeiana (Shaw, 1802) |
amebul* |
Anthrophony |
Airplane |
airplane* |
|
Chainsaw |
chainsaw* |
||
Gunshot |
gunshot* |
||
Sensor noise |
sensor* |
||
Truck beep |
truck* |
||
Vehicle |
vehicle* |
||
Bird |
American Goldfinch |
Spinus tristis (Linnaeus, 1758) |
amegfi |
American Robin |
Turdus migratorius (Linnaeus, 1766) |
amerob |
|
Band-tailed Pigeon |
Patagioenas fasciata (Say, 1822) |
batpig1 |
|
Bewick's Wren |
Thryomanes bewickii (Audubon, 1827) |
bewwre |
|
Black-capped Chickadee |
Poecile atricapillus (Linnaeus, 1766) |
bkcchi |
|
Black-headed Grosbeak |
Pheucticus melanocephalus (Swainson, 1827) |
bkhgro |
|
Black-throated Gray Warbler |
Setophaga nigrescens (Townsend, 1837) |
btywar |
|
Brown Creeper |
Certhia americana (Bonaparte, 1838) |
brncre |
|
Bushtit |
Psaltriparus minimus (Townsend, 1837) |
bushti |
|
Canada Jay |
Perisoreus canadensis (Linnaeus, 1766) |
gryjay |
|
Cassin's Vireo |
Vireo cassinii (Xantus de Vesey, 1858) |
casvir |
|
Cedar Waxwing |
Bombycilla cedrorum (Vieillot, 1808) |
cedwax |
|
Chestnut-backed Chickadee |
Poecile rufescens (Townsend, 1837) |
chbchi |
|
Chipping Sparrow |
Spizella passerina (Bechstein, 1798) |
chispa |
|
Common Raven |
Corvus corax (Linnaeus, 1758) |
comrav |
|
Cooper's Hawk |
Accipiter cooperii (Bonaparte, 1828) |
coohaw |
|
Dark-eyed Junco |
Junco hyemalis (Linnaeus, 1758) |
daejun |
|
Downy Woodpecker |
Dryobates pubescens (Linnaeus, 1766) |
dowwoo |
|
Dusky Flycatcher |
Empidonax oberholseri (Phillips, 1939) |
dusfly |
|
Evening Grosbeak |
Hesperiphona vespertina (Cooper, 1825) |
evegro |
|
Golden-crowned Kinglet |
Regulus satrapa (Lichtenstein, 1823) |
gockin |
|
Hairy Woodpecker |
Leuconotopicus villosus (Linnaeus, 1766) |
haiwoo |
|
Hammond's Flycatcher |
Empidonax hammondii (Xantus de Vesey, 1858) |
hamfly |
|
Hermit Thrush |
Catharus guttatus (Pallus, 1811) |
herthr |
|
Hermit Warbler |
Setophaga occidentalis (Townsend, 1837) |
herwar |
|
Hutton's Vireo |
Vireo huttoni (Cassin, 1851) |
hutvir |
|
Lazuli Bunting |
Passerina amoena (Say, 1822) |
lazbun |
|
MacGillivray's Warbler |
Geothlypis tolmiei (Townsend, 1839) |
macwar |
|
Mountain Chickadee |
Poecile gambeli (Ridgway, 1886) |
mouchi |
|
Mountain Quail |
Oreortyx pictus (Douglas, 1829) |
mouqua |
|
Nashville Warbler |
Leiothlypis ruficapilla (Wilson, 1811) |
naswar |
|
Northern Flicker |
Colaptes auratus (Linnaeus, 1758) |
norfli |
|
Northern Pygmy-Owl |
Glaucidium gnoma (Wagler, 1832) |
nopowl |
|
Olive-sided Flycatcher |
Contopus cooperi (Nuttall, 1831) |
olsfly |
|
Orange-crowned Warbler |
Leiothlypis celata (Say, 1822) |
orcwar |
|
Pacific Wren |
Troglodytes pacificus (Baird, 1864) |
pacwre1 |
|
Pacific-slope Flycatcher |
Empidonax difficilis (Baird, 1858) |
pasfly |
|
Pileated Woodpecker |
Dryocopus pileatus (Linnaeus, 1758) |
pilwoo |
|
Pine Siskin |
Spinus pinus (Wilson, 1810) |
pinsis |
|
Purple Finch |
Haemorhous purpureus (Gmelin, 1789) |
purfin |
|
Red-breasted Nuthatch |
Sitta canadensis (Linnaeus, 1766) |
rebnut |
|
Rooster (Red Junglefowl) |
Gallus gallus (Linnaeus, 1758) |
redjun1 |
|
Rufous Hummingbird |
Selasphorus rufus (Gmelin, 1788) |
rufhum |
|
Say's Phoebe |
Sayornis saya (Bonaparte, 1825) |
saypho |
|
Sooty Grouse |
Dendragapus fuliginosus (Ridgway, 1873) |
soogro1 |
|
Spotted Towhee |
Pipilo maculatus (Swainson, 1827) |
spotow |
|
Steller's Jay |
Cyanocitta stelleri (Gmelin, 1788) |
stejay |
|
Swainson's Thrush |
Catharus ustulatus (Nuttall, 1840) |
swathr |
|
Townsend's Solitaire |
Myadestes townsendi (Audubon, 1838) |
towsol |
|
Townsend's Warbler |
Setophaga townsendi (Townsend, 1837) |
towwar |
|
Varied Thrush |
Ixoreus naevius (Gmelin, 1789) |
varthr |
|
Warbling Vireo |
Vireo gilvus (Vieillot, 1808) |
warvir |
|
Western Tanager |
Piranga ludoviciana (Wilson, 1811) |
westan |
|
Western Wood-Pewee |
Contopus sordidulus (Sclater, 1859) |
wewpew |
|
Wild Turkey |
Meleagris gallopavo (Linnaeus, 1758) |
wiltur |
|
Wilson's Warbler |
Cardellina pusilla (Wilson, 1811) |
wlswar |
|
Wrentit |
Chamaea fasciata (Gambel, 1845) |
wrenti |
|
Yellow-rumped Warbler |
Setophaga coronata (Linnaeus, 1766) |
yerwar |
|
Geophony |
Rain |
rain* |
|
Mammal |
Dog |
Canis lupus familiaris (Linnaeus, 1758) |
dog* |
Douglas squirrel |
Tamiasciurus douglasii (Bachman, 1839) |
dousqu* |
The audio clips comprising this dataset were recorded during the initial hour following sunrise, spanning the time frame from 01-05-2022 to 25-09-2022. However, due to variations in the spatial distribution of our recording units and the effects of our filtering criteria, recordings from May are relatively over-represented and recordings from California only occurred during May and June (Table
Creative Commons Attribution (CC-BY) 4.0 License
This dataset describes the acoustic recordings included in this dataset. The acoustic recordings described in the dataset are available through an online data repository DOI: https://zenodo.org/doi/10.5281/zenodo.8047849.
Column label | Column description |
---|---|
site | Site name. |
replicate | An ordinal label indicating the random draw label: ‘A’, ‘B’, or ‘C’. |
recording_date | Recording date and time formatted as “Year-Month-Day Hour:Minute:Second”. |
annotated | Categorical assignment describing whether a recording was completely annotated: ‘complete,’ ‘partial,’ or ‘not annotated’. |
file | Wav file name. |
zip_file | The zip file location of the file. |
This dataset lists all annotations from the fully annotated recordings.
Column label | Column description |
---|---|
file | Wav file name. |
start | Start time of the 2-second clip in seconds. |
end | End time of the 2-second clip in seconds. |
eBird_2021 | 2021 species identification eBird code. |
label | Sonotype label concatenates the 2021 eBird taxonomy code and the sound type label. |
This dataset lists all annotations from the partially annotated recordings.
Column label | Column description |
---|---|
file | Wav file name. |
start | Start time of the 2-second clip in seconds. |
end | End time of the 2-second clip in seconds. |
clip_complete | Binary indicator for whether the clip was completely labelled. |
eBird_2021 | 2021 species identification eBird code. |
label | Sonotype label comprising a concatenation of the 2021 eBird taxonomy code and the sound type label. |
This dataset describes the focal acoustic sounds included in the recording annotations.
Column label | Column description |
---|---|
label | Sonotype label comprising a concatenation of the 2021 eBird taxonomy code and the sound type label. |
eBird_2021 | 2021 eBird taxonomy species_code. |
sound | Sound type label. |
common_name | The common name of the sound source. For avian species, the scientific name follows Clement’s taxonomy outlined in the 2021 eBird taxonomy. |
scientific_name | The scientific name of the biotic sound source. For avian species, the scientific name follows Clement’s taxonomy outlined in the 2021 eBird taxonomy. |
taxonomic_authority | Primary taxonomic authority. |
description | Biological and phonetic description of the target sound. |
n_files | Total number of audio files containing at least one of the target labels. |
n_annotations | Total number of label-specific annotations in the fully annotated data. |
This dataset lists the environmental characteristics at each recording station. Units of measurements for appropriate covariates are in parentheses.
Column label | Column description |
---|---|
site | Site name. |
replicate | An ordinal label indicating whether the row describes a random sample ‘A’, ‘B’ or ‘C’. |
state | State location of survey site. |
township_range | Township and range identifier of the survey site. The township was data obtained from three sources: CA, OR, WA. |
age_dom_2017 | Basal area weighted stand age, based on dominant and codominant trees (years). |
ba_ge_3_2017 | Basal area of live trees >= 2.5 cm dbh (m2/ha). |
bac_ge_3_2017 | Basal area of live conifers >= 2.5 cm dbh (m2/ha). |
bah_ge_3_2017 | Basal area of live hardwoods >= 2.5cm dbh (m2/ha). |
bph_ge_3_crm_2017 | Component Ratio Method biomass of all live trees >= 2.5 cm (kg/ha). |
bphc_ge_3_crm_2017 | Component Ratio Method biomass of all live conifers >= 2.5 cm (kg/ha). |
bphh_ge_3_crm_2017 | Component Ratio Method biomass of all live hardwoods >= 2.5 cm (kg/ha). |
cancov_2017 | Canopy cover of all live trees (percent). |
cancov_con_2017 | Canopy cover of all conifers (percent). |
cancov_hdw_2017 | Canopy cover of all hardwoods (percent). |
cancov_layers_2017 | Number of tree canopy layers present (number of layers). |
conplba_2017 | Conifer tree species with the plurality of basal area (raster to alphanumeric look-up table available at source). |
covcl_2017 | Cover class based on cancov (raster to alphanumeric look-up table available at source). |
ddi_2017 | Diameter diversity index |
fortypba_2017 | Forest type, which describes the dominant tree species of current vegetation (raster to alphanumeric look-up table available at source). |
hdwplba_2017 | Hardwood tree species with the plurality of basal area (raster to alphanumeric look-up table available at source). |
mndbhba_2017 | Basal-area weighted mean diameter of all live trees (cm). |
mndbhba_con_2017 | Basal-area weighted mean diameter of all live conifers (cm). |
mndbhba_hdw_2017 | Basal-area weighted mean diameter of all live hardwoods (cm). |
qmd_dom_2017 | The quadratic mean diameter of all dominant and codominant trees (cm). |
qmd_ht25_2017 | The quadratic mean diameter in inches of trees whose heights are in the top 25% of all tree heights (cm). |
qmdc_dom_2017 | The quadratic mean diameter of all dominant and codominant conifers (cm). |
qmdh_dom_2017 | The quadratic mean diameter of all dominant and codominant hardwoods (cm). |
sbph_ge_25_2017 | Biomass of snags >= 25 cm dbh and >= 2m tall (lb). |
sdi_reineke_2017 | Reineke's stand density index. |
sizecl_2017 | Size class, based on qmd_dom and cancov (raster to alphanumeric look-up table available at source). |
stndhgt_2017 | Stand height, computed as the average height of all dominant and codominant trees (m). |
stph_ge_25_2017 | Density of snags >= 25 cm dbh and >= 2 m tall (trees/ha). |
struccond_2017 | Structural condition (raster to alphanumeric look-up table available at source). |
svph_ge_25_2017 | Volume of snags >= 25 cm dbh and >= 2 m tall (m3/ha). |
tph_ge_3_2017 | The density of live trees >= 2.5 cm dbh (trees/ha). |
tphc_ge_3_2017 | The density of live conifers >= 2.5 cm dbh (trees/ha). |
tphh_ge_3_2017 | The density of live hardwoods >= 2.5 cm dbh (trees/ha). |
treeplba_2017 | Tree species with the plurality of basal area (raster to alphanumeric look-up table available at source). |
vegclass_2017 | Vegetation class based on cancov, bah_prop, qmd_dom (raster to alphanumeric look-up table available at source). |
vph_ge_3_2017 | The volume of live trees >= 2.5 cm dbh (m3/ha). |
vphc_ge_3_2017 | The volume of live conifers >= 2.5 cm dbh (m3/ha). |
vphh_ge_3_2017 | The volume of live hardwoods >= 2.5 cm dbh (m3/ha). |
dem_30m | Digital elevation model at 30 m² resolution (m). |
northness_30m | A cosine transformation of aspect to demonstrate the orientation of a land relative to a north-facing land derived from dem_30m. |
slope_30m | Estimate of land slope at 30 m² resolution derived from dem_30m. |
tpi5x5_30m | Mean difference of the central point to a focal square of the surrounding 5 × 5 grid cells derived from dem_30m. |
vrm_30m | Variation in slope and aspect derived from dem_30m. |
an_precip_1km | Average precipitation at a 1 km2 resolution averaged from 1970-2000 (mm). |
minT_1km | Average minimum temperature at a 1 km2 resolution averaged from 1970-2000 (degrees Celcius). |
maxT_1km | Average maximum temperature at a 1 km2 resolution averaged from 1970-2000 (degrees Celcius). |
This dataset describes the environmental characteristics included in environmental_characteristics.
Column label | Column description |
---|---|
covariate | Covariate name. |
type | Value type of variable. |
range | The range of values extracted across our survey sites. The values in this cell represent the value minimum to the value maximum. |
unit | A description of the variable units of measurement. |
description | A description of the variable, including a brief discussion of the methods used to create the variable. |
source | Variable source. |
This dataset describes the annotator identification and annotation method for each 2-second window.
Column label | Column description |
---|---|
file | Wav file name. |
start | Start time of the 2-second clip in seconds. |
end | End time of the 2-second clip in seconds. |
method | The annotation method used to label the 2-second clip. This label is only available for a subset of clips used to estimate annotation speed. |
annotator | The annotator identifier for the 2-second clip. |
The fully annotated acoustic recordings are available for download in a zip file of uncompressed wav format files.
The partial and unannotated recordings are available in 11 zip files of uncompressed wav format files.
Descriptive data dictionaries are available for download as a pdf file.
We fully annotated 11.75 hours of audio with 32,994 labels for 115 sonotypes. An additional 216 files were partially annotated with 5,278 labels for 53 sonotypes. We also provide 20,737 auditing labels indicating clip-level completion status. The most frequently annotated species were Red-breasted Nuthatch Sitta canadensis (Linnaeus, 1766; eBird code: rebnut; n annotations = 2,496), Pacific Wren Troglodytes pacificus (Baird, 1864; eBird code: pacwre1; n annotations = 2,259), Hermit Thrush (eBird code: herthr; n annotations = 1,750), Swainson’s Thrush Catharus ustulatus (Nuttall, 1840; eBird code: swathr; n annotations = 1,519), Pacific-slope Flycatcher Empidonax difficilis (Baird, 1858; eBird code: pasfly; n annotations = 1,405) and Golden-crowned Kinglet Regulus satrapa (Lichtenstein, 1823; eBird code: gockin; n annotations = 1,368; Fig.
This vertical barplot visualises the frequency of annotations for the most prevalent species within the annotated dataset. The y-axis lists species by their 2021 eBird codes, ordered from most to least frequent (see Table 1 for common names). The x-axis displays the cumulative annotation count for each species. More prevalent species occur towards the bottom and have higher annotation counts. The plot reveals that a few common species dominate annotations, while many are annotated infrequently.
We annotated an average of 695 windows per hour (σ = 363). However, the annotation rate varied between annotators and methods. The model-assisted method appeared to increase the rate for both annotators relative to the linear method (Table
Summary statistics for the annotation rate, measured in windows annotated per hour, from two annotators employing both linear and model-assisted annotation protocols. The summary includes the following metrics: mean, standard deviation (sd), minimum rate (min.) and maximum rate (max.).
Annotation method |
Annotator |
mean |
sd |
min. |
max. |
linear |
1 |
651.6 |
274.5 |
333.3 |
1125.0 |
2 |
580.3 |
605.3 |
160.7 |
2250.0 |
|
model-assist |
1 |
883.8 |
292.2 |
303.2 |
1582.4 |
2 |
496.3 |
227.0 |
201.4 |
1170.7 |
Two of the aggregated biophonic sound groups (Parulidae complex: paruli, Setophaga complex: setoph) consisted of groups of similar sound types (eBird codes for sound classes included in the Parulidae complex: macwar, naswar, wlswar, yerwar; eBird codes for sound classes included in the Setophaga complex: btywar, herwar, towwar) that we were unable to assign to a species-level eBird code confidently. Another aggregated biophonic sound (Unknown chip: unk) consisted of unknown avian chip calls, which we were also unable to assign to a species-level eBird code confidently. We could not confidently differentiate six biotic sound groups. To gain insight into the acoustic structure of these groups, we used t-distributed stochastic neighbour embedding (t-SNE) of BirdNET embeddings (
Two-dimensional t-SNE (t-distributed stochastic neighbour embedding) plots of the BirdNET embeddings for two aggregated biotic classes and unambiguous examples from individual species included in the aggregated classes. Each data point on the plot corresponds to an individual 2-second audio clip. Panel A plots the t-SNE embedding for the paruli aggregated class, which includes MacGillivray's Warbler Geothlypis tolmiei (Townsend, 1839; eBird code: macwar), Nashville Warbler Leiothlypis ruficapilla (Wilson, 1811; eBird code: naswar), Yellow-rumped Warbler Setophaga coronata (Linnaeus, 1766; eBird code: yerwar) and Wilson's Warbler Cardellina pusilla (Wilson, 1811; eBird code: wlswar). Panel B plots the t-SNE embedding for the Setophaga aggregated class, which includes Hermit Warbler Setophaga occidentalis (Townsend, 1837; eBird code: herwar), Townsend's Warbler Setophaga townsendi (Townsend, 1837; eBird code: towwar) and Black-throated Gray Warbler Setophaga nigrescens (Townsend, 1837; eBird code: btywar). This visualisation compares the aggregated classes to known examples from key species, evaluating the overlap of individual species embeddings relative to their assigned aggregated class.
Two-dimensional t-SNE (t-distributed stochastic neighbour embedding) plots of the birdnet embeddings for the aggregated biotic unknown avian chip vocalisation class (eBird code: unk). Each data point on the plot corresponds to an individual 2-second audio clip. Panels B, C and D provide detailed spectrograms for selected audio clips marked by opaque black points on the t-SNE plot. Panel B exemplifies a typical audio clip near the centre of the primary unknown chip cluster within the t-SNE plot. Many audio clips in this cluster contain only an avian chip vocalisation. Panel C features the spectrogram of an audio clip from the most negative sub-cluster along the t-SNE axis 2. Audio clips within this sub-cluster primarily contain vocalisations from the aggregated Setophaga class. Panel D displays the spectrogram of an audio clip from the most positive cluster along t-SNE axis 1. Audio clips within this cluster predominantly consist of Red-breasted Nuthatch vocalisations.
In its essence, a labelled acoustic dataset is a presence-absence dataset. When we pair species-level labels with local environmental characteristics, we can explore the relative presence of species across environmental gradients. For example, Varied Thrush Ixoreus naevius (Gmelin, 1789) and Pacific Wren prefer older forests, implying that their likelihood of occurrence within such habitats should be higher when compared to the baseline sampling rate of older forests and, for any given covariate, a species with a typical response pattern should closely align with that baseline sampling rate (Fig.
Kernel density plots of species occurrence across gradients of Basal area weighted stand age (bandwidth = 2000), Reineke’s stand density index (bandwidth = 500) and canopy cover of conifer trees (bandwidth = 5000). The species-specific probability densities are shown relative to the base rate of sampling occurrence across each environmental gradient. Specialist species with respect to an environmental gradient should show higher or lower probabilities relative to the sampling base rate within some range of the environmental gradient (i.e. > or < 0), whereas generalist species for a given environmental gradient should match the sampling base rate of occurrence (i.e. ~ 0).
Recent advances in computational algorithms have made passive acoustic monitoring more accessible (
Acoustic data collection was funded and collected by the US Forest Service and the US Bureau of Land Management. Annotation work was funded by Google. We thank the many biologists who collected and processed the data compiled here. The use of trade or firm names in this publication is for reader information and does not imply endorsement by the U.S. Government of any product or service.
M.J.W., T.D. and D.B.L. conceived of the analysis. D.B.L. deployed and maintained PAM recorders. J.M.A.J. and Z.R. managed annual data management. M.J.W., J.M.A.J. and Z.R. collated this dataset. M.J.W., T.D., A.B.F. and J.T. developed the annotation protocol. A.B.F. and J.T. annotated the acoustic recordings. M.J.W. and R.S.S. collected and validated spatial data extractions. M.J.W. and A.B.F. analysed the data. M.J.W. led the writing of the manuscript. All authors contributed critically to the drafts and gave final approval for publication.