- Researchers have developed machine learning techniques to identify bird song from thousands of hours of field recordings, using the information to uncover variations in migratory songbirds’ arrival to their Arctic breeding grounds.
- They deployed automated listening devices during spring over five years, analyzed vocal activity to estimate when birds arrived at their breeding sites, and assessed relationships between vocal activity and environmental conditions.
- They found that the acoustically derived estimates of the birds’ arrival dates were similar to those determined using standard field surveys.
- Temperature and presence of snow affected the birds’ calling patterns, suggesting that collecting corresponding weather data could help avoid bias in using acoustic monitoring to assess population dynamics.
It’s June, and migratory songbirds in the northern hemisphere are at their summer breeding grounds, having traveled thousands of miles from their warm-weather overwintering areas.
Birds migrate as far north as the Arctic to take advantage of its large but short-lived surge in insect food and its few predators. The timing of their arrival is critical because their breeding cycles must match seasonal food availability for their chicks to survive.
Scientists have shown that as spring temperatures rise, many bird populations are, in fact, migrating north and arriving earlier in the season at their breeding sites, where climate-related shifts in breeding-ground conditions, including environmental conditions and food availability, may help or hinder reproduction of individual species.
Most songbirds are too small to carry GPS tracking tags scientists would typically use to follow their migrations north, but they do call intensely once they arrive there in preparation for breeding.
To study trends in migration timing, scientists have begun setting out microphones to listen for particular species or the bird diversity at specific sites. Placing numerous relatively inexpensive acoustic listening devices in the field allows researchers to better monitor wildlife communities in remote places and across larger scales than field surveys typically can.
A multi-institutional research team deployed automated listening devices over five spring breeding seasons at sites in Alaska to capture the vocalizations of two common breeding songbird species. White-crowned sparrows (Zonotrichia leucophrys) and Lapland longspurs (Calcarius lapponicus) both fly to these sites each spring to mate and raise their young.
Autonomous recorders in the field can collect data 24/7, and their use has relied on trained experts to listen to the recordings and detect a target species or tally the species present at a given site. However, automated recordings of whole bird communities over hours or days produce data sets too large to review manually.
Automating analysis of birdsong patterns
To facilitate the use of acoustic devices in studying whole communities of breeding birds, the researchers developed automated signal processing and machine learning algorithms adapted from human speech research to estimate from acoustic signals when songbird communities arrived each spring at four breeding sites in Alaska. For five seasons (2010 to 2014), the research team recorded songbird vocalizations at the sites at regular intervals from early May through July.
They developed and trained a supervised machine learning algorithm, one that includes human input, to pick out calls of target songbirds from thousands of hours of field recordings that also contained noise from trucks, wind, rain, mosquitos and other bird species. They used the call data to produce a daily community Vocal Activity Index (VAI), a relative measure of the abundance of bird vocalizations at each site. They analyzed the daily VAI values to estimate the dates that the bulk of these birds had arrived at their breeding sites and any relationships between the VAI and environmental conditions, including temperature, wind and snow cover.
The researchers also analyzed the sound data using an unsupervised classification, which does not use listener input but classifies data into groups that represent like items, to see if it could pick out the bird songs on its own and use them to estimate the arrival date.
Cool birds don’t sing
The researchers found that songbird vocal activity varied both in time (days, weeks and years) and the surrounding environment.
They state in their paper, “We found that daily fluctuations in snow cover, air temperature, wind speed, atmospheric pressure, and precipitation had a significant impact on the VAI and explained a large proportion of variance.”
In particular, they found calling activity increased noticeably on snow-free days, and they suggest that birds rely on snow-free patches of tundra for food and shelter. Singing takes energy, even more so on colder days; the songbirds in this study either moved on or remained quiet during unfavorable weather.
They also found that both the supervised and the unsupervised arrival date estimates closely approximated what human observers recorded at the sites.
Tools to study a range of calling creatures
The success of the automated analyses is good news for researchers studying animal movements and population dynamics and could help scientists better understand patterns of migration and how they may be adapting to changes in climate patterns.
“These tools could speed up the analysis of acoustic datasets packed with biodiversity information valuable to conservationists and others,” Andrew Farnsworth, a researcher at the Cornell Lab of Ornithology, who was not involved in the study, said in a statement. “Understanding the dynamics of songbird arrival and breeding timing is the doorway to thinking about climate change and how temperature, weather and snowfall are affecting various species.”
Listen to a Gambel’s white-crowned sparrow singing near Toolik Field Station in arctic Alaska. Audio credit: Oliver et al. 2018
A functional unsupervised machine learning method could potentially be extended to any dataset of animal vocalizations. An unsupervised automated analysis does not need to be trained with a reference database of calls, as it does not need to be told what it is hearing.
“Our methods could be retooled to detect the arrival of birds and other vocal animals in highly seasonal habitats,” said the study’s lead author, Ruth Oliver, a graduate student at Columbia University. “This could allow us to track large-scale changes in how animals are responding to climate change.”
The study also showed that acoustic monitoring must consider environmental factors, such as temperature, that may influence how much animals call and thus lead to biased conclusions, as the listening devices cannot distinguish silence from absence.
“Our findings demonstrate that the correct interpretation of avian vocal activity to estimate relative songbird abundance requires pairing of acoustic data collection with meteorological data, as well as consideration of the study communities’ breeding phenology [breeding stage].”
“It’s still unclear how songbirds will cope if spring comes even earlier or later than it did during our study period,” said co-author Natalie Boelman. “Species also time their migration and breeding with day length, which isn’t shifting with climate change. Species whose migratory response is hard-wired to day length alone may not adapt as well to a changing environment.”
Citation
Oliver, R., Ellis, D., Chmura, H., Krause, J.S., Pérez, J.H., Sweet, K.S., Gough, L., S. K., Wingfield, & J. Boelman, N.. (2018). Eavesdropping on the Arctic: Automated bioacoustics reveal dynamics in songbird breeding phenology. Science Advances 20 Jun 2018: Vol. 4, no. 6, eaaq1084 DOI: 10.1126/sciadv.aaq1084
FEEDBACK: Use this form to send a message to the editor of this post. If you want to post a public comment, you can do that at the bottom of the page.
Credits
