Passive Acoustic Monitoring |
|
In 2019 Dr. Lesmeister started the PNW Bioacoustics Lab to leverage emerging technologies to non-invasively study wildlife and monitor biodiversity at large scales. We are using passive acoustic monitoring as the primary means to study northern spotted owls in a non-invasive manner. This is done with autonomous recording units, or ARUs, that record owls calling in the wild. These units have been deployed by the Lesmeister Lab and PNW Bioacoustics Lab since 2017, and over 4,000 sites are monitored annually throughout federally managed forest lands in the Pacific Northwest.
|
Millions of hours of these acoustic data are analyzed through a deep convolutional neural network, PNW-Cnet. This network effectively automates the detection of vocal activity in target species. Once detections are confirmed by lab technicians, we use these data to generate encounter histories for occupancy analyses for spotted owls and several other species. We have fully transitioned the monitoring of northern spotted owls in the Pacific Northwest away from mark and re-sight demography to passive acoustic monitoring. We have published several papers that have helped to inform the best transition strategy and the future of the monitoring program.
|
|
|
|
|
WHAT WE'VE DISCOVERED
Passive acoustic monitoring minimizes human interaction in the complex relationship between spotted owls and their main competitor, the barred owl. The Lesmeister Lab switched almost exclusively to bioacoustic data collection in 2021 as per the guidance of the Northwest Forest Plan Interagency Regional Monitoring Plan. A focus on bioacoustics does shift research objectives: population size and distribution become the key monitoring indicators, while survival, recruitment, and population growth estimates gathered through traditional demography work are no longer available. While there are drawbacks to ending our role in demographic research, a future in bioacoustics provides us with the ability to monitor more of the forested landscape and to gather data on multiple species simultaneously.
Photo by Todd Sonflieth / OPB
|
The PNW Bioacoustics Lab and the Lesmeister Lab worked to generate the best available science to inform the transition in order to ensure data on northern spotted owls continued to be gathered in the most accurate manner possible. Former lab member Lelia Duchac, seen deploying an ARU on the left, conducted critical research in 2017 for her MS thesis that illustrated the viability of passive acoustic monitoring for spotted owls and barred owls. Her work continues to help design our current research methodologies. She also used passive acoustics to study post-fire landscape use by forest owls.
|
We survey eight study sites in Washington, Oregon, and northern California. Each site is overlaid with a 5-km² hexagonal grid and 20% of these hexagons are chosen at random to host autonomous recording units (ARUs). Each chosen hexagon has four stations where research technicians deploy ARUs for a total of six weeks with sound recorded at specific intervals throughout the day. Although hexagons include a matrix of public and private forest land, ARU stations are only placed on publicly owned land.
In addition to the eight study sites, ARUs are also deployed across a 2% sample of federal forest lands throughout the study area.
In addition to the eight study sites, ARUs are also deployed across a 2% sample of federal forest lands throughout the study area.
PROVEN TO WORKARUs can effectively detect northern spotted owls when they are present, even on a landscape with high barred owl density. Our research methodology was tested and proven by Lelia Duchac and other lab members, with a final publication released in 2020. We are confident in our ability to continue to monitor the threatened northern spotted owl utilizing bioacoustic technology.
THE PUBLICATIONRecordings collected between March and July of 2017 at three Oregon and Washington sites were analyzed using our convolutional neural network (CNN). ARUs were shown to effectively detect northern spotted owl vocalizations, even in areas with many barred owls, at a rate of 95% when the units recorded for three weeks. This data illustrated spatial and temporal patterns to owl vocalization as well as differences in vocal intensity, allowing researchers to distinguish nesting pairs from non-territorial birds. Data collected by ARUs will serve as a valuable tool in determining nesting status and gathering other demographic information in location where owls are known to be residing.
|
ENTIRE STUDY AREA WITH HEXAGONS
|
The deep convolutional neural network, PNW-Cnet, has been in development over several years by the Lesmeister Lab in partnership with computer scientists at Oregon State University and worked on extensively by lab member Zack Ruff. It is trained to detect 80 species and over a dozen human sounds. Running data through PNW-Cnet reduces human effort by over 99%. It is an invaluable tool that allows for a quick turnaround time between data collection and processing, shortening the timeline of analysis.
Under current protocol, audio data gathered by ARUs in the field are converted to spectrogram images. These images are run through the PNW-Cnet to be classified by species and call type. Clips of each identifiable vocalization are sent to research assistants for review, and those verified calls generate encounter histories for each target species. Encounter histories inform occupancy analysis and will drive population estimates for northern spotted owls as well as additional target species, including the barred owl.
PNW-CNET TRAINING AND TESTING
In order to get to this streamlined workflow, extensive training of the PNW-Cnet occurred behind the scenes to ensure its capacity to correctly distinguish and classify species vocalizations. The PNW-Cnet model was compiled and trained in Python using Keras, an open-source, machine learning-focused application programming interface to Google’s TensorFlow software library. Extensive training and testing ran hundreds of thousands of spectrogram clips through the PNW-Cnet, and yielded high results for detection of northern spotted owls. Specificity and sensitivity of detection for other species of owl was similarly high. However, these rates were lower for other bird species and mammals vocalizations, as seen in the F1 score graph below which is a gage of overall model performance.
|
Using our trained CNN, precision of detecting the northern spotted owls will exceed 99%. Details on training and working with the PNW-Cnet were published in 2021 by lab members Zach Ruff, Damon Lesmeister, Cara Appel, and OSU’s Christopher Sullivan.
Our initial research training and testing the PNW-Cnet was published in 2019. |
|