Quantifying bird communities using automated soundscape recordings

• Katherine Kaufmann, Wildlife Ecology and Conservation, University of Delaware

• Jeff Buler, Entomology and Wildlife Ecology, University of Delaware

Many North American bird species favor or even require early successional habitat to complete their life cycles. However, increased development and a lack of indigenous land management practices like prescribed burns have reduced the amount of usable early successional habitat. At the Letterkenny Army Depot in Franklin County, PA, the Wild Northern Bobwhite Quail Restoration project has implemented management strategies such as mechanical plant removal, herbicide application, and fire since 2017 to increase early successional habitat for Northern Bobwhite Quail. The new viable habitat benefits many other birds, as well. My objective was to quantify seasonal variation in the bird community across different habitat and management types at the site. We deployed 21 Autonomous Recording Units (ARUs) in 7 different habitat types. The ARUs recorded birds from one hour before sunrise to two hours after sunset for three days per month from 2024 to 2026. We used the Cornell Lab of Ornithology’s BirdNET software to detect and classify bird species from the recordings via an artificial neural network. For each species, I manually reviewed a sample of recordings with varying BirdNET confidence scores and identified true and false detections, calculating 0.99, 0.95, 0.90, and 0.85 confidence thresholds for true positives. Field Sparrow, Eastern Whip-poor-will, Eastern Towhee, Grasshopper Sparrow, and Gray Catbird had the most detections in the breeding season (March-August); and White-crowned Sparrow, White-throated Sparrow, Field Sparrow, American Woodcock, and Gray Catbird had the most detections in the nonbreeding season (September-February). In 38 proofed species, 0.95 confidence thresholds ranged from -0.62 to 0.77. Northern Bobwhite, the focal species of the restoration project, achieves a 95% confidence threshold at a BirdNET confidence score of 0.44.