Automated Detection of Bird Roosts Using NEXRAD Radar Data and Convolutional Neural Networks

Abstract

NEXRAD radars have proven to be an effective tool for detecting bird roosts for several species or birds, however manually locating these roosts in radar images is a time consuming process. We introduce a Convolutional Neural Network trained to automatically determine whether each individual radar image contains at least one Purple Martin or Tree Swallow roost. Radars give us a continental-scale snapshot of an entire vertebrate population. Many fields within ecology conservation could benefit from automated detection of bird roosts, and we are able to find bird roosts for species that are visible in radar imagery with 90 percent accuracy. We use a dataset of radar images that contain Purple Martin roosts and Tree Swallow roosts in the Eastern half of the United States. We show that Convolutional Neural Networks (CNNs) are an effective method for automating the bird roost detection. CNNs have recently revolutionized image classification largely because CNNs capture spatial components of images. We hypothesized that these same principles can be applied to radar data. To further improve the accuracy of bird roost detection, machine learning techniques such as batch normalization and transfer learning are applied to the CNN. Our results show that CNNs are a promising approach for bird roost detection for legacy radar data and dual polarization radar data.