Modeling Swallow Roosts Using Weather Radar
Date
2017-06-30
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Every day, large flocks of swallows fly away from nocturnal roosting locations at
sunrise. These flocks, comprised of any one of the 4-5 swallow species present
in the US, are known as roosts. Along with storm systems and precipitation,
roosts are visible on weather radar. It is estimated over 200 million radar scans
have been collected containing potentially invaluable information pertaining to
bird roosting behavior. Biologists could use this data to gain a better understanding
of bird migratory patterns, and the effect of climate change on avian
behavior. However, it is difficult for biologists to interpret information from the
radar scans due to the lack of established methods for detecting roosts in these
scans.
This thesis project seeks to simplify the process of interpreting bird roost
radar data by developing an automated method of detecting and locating roosts.
We consider two approaches to solving this problem; the development of empirical
parametric models based on data products returned by the radar, and
adaptation of existing computer vision algorithms. For the first approach, we
find a hypothesized Gaussian relationship between reflectivity of volumes of atmosphere,
and their distance from the center of a labeled roost. We then use
this relationship to construct a model for localizing roosts within radar scans,
and compare the results to a similar model that uses radial velocity. For the
second approach, we begin by adapting an open-source Haar feature-based cascade
classifier object detection method. We also consider convolutional neural
networks.
In each case, we observe the suitability of the model for detection or localization
of roosts within radar scans, as well as precision and accuracy compared
to human annotation. Our experiments on a subset of a pre-labeled database
of roost scans show that a machine learning approach to roost labeling can
improve on existing methods of annotation, though false detections due to the
difficulty of distinguishing between roosts and precipitation are an ongoing issue.
Fine-tuning of the models could improve precision.
Description
Keywords
Computer Science, Computational Ecology, Swallow, Roost, Machine Learning, Swallow Roost