7A.6 Locating Bird Roosts Using NEXRAD Radar Data and Image Segmentation

Wednesday, 15 January 2020: 9:45 AM
156BC (Boston Convention and Exhibition Center)
Katherine Avery, Univ. of Oklahoma, Norman, OK; and A. McGovern, E. Bridge, and J. F. Kelly

NEXRAD radars can effectively detect flying animals, such as groups of birds, bats, and insects. Further, these radars are demonstrably useful for detecting the existence of bird roosting locations, which appear on radar as distinctive rings of high reflectivity. Detecting and locating bird roosts have a variety of applications from ecological conservation to wind farm placement and air traffic control. Convolutional neural networks (CNNs) have proven effective for detecting the presence of roosts in NEXRAD images. After training on a dataset of purple martin and tree swallow roosts in the eastern half of the U.S, the network makes correct determinations above 90 percent of the time and has an area under curve (AUC) above 0.9. However, determining the exact locations, rather than just the presence, of bird roosts is more useful. We apply deep learning and image segmentation, which show promising results for determining roost location. To increase performance, we augment our dataset 32 times its original size through a variety of image transformation. We present the preliminary results of identifying the specific locations of the roosts in this augmented set.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner