Meteor Head Echo Detection at Multiple High-Power Large-Aperture Radar Facilities via a Convolutional Neural Network Trained on Synthetic Radar Data.

High-power large-aperture radar instruments are capable of detecting thousands of meteor head echoes within hours of observation, and manually identifying every head echo is prohibitively time-consuming. Previous work has demonstrated that convolutional neural networks (CNNs) accurately detect head echoes, but training a CNN requires thousands of head echo examples manually identified at the same facility and with similar experiment parameters. Since pre-labeled data is often unavailable, a method is developed to simulate head echo observations at any given frequency and pulse code. Real instances of radar clutter, noise, or ionospheric phenomena such as the equatorial electrojet are additively combined with synthetic head echo examples. This enables the CNN to differentiate between head echoes and other phenomena. CNNs are trained using tens of thousands of simulated head echoes at each of three radar facilities, where concurrent meteor observations were performed in October 2019. Each CNN is tested on a subset of actual data containing hundreds of head echoes, and demonstrates greater than 97% classification accuracy at each facility. The CNNs are capable of identifying a comprehensive set of head echoes, with over 70% sensitivity at all three facilities, including when the equatorial electrojet is present. The CNN demonstrates greater sensitivity to head echoes with higher signal strength, but still detects more than half of head echoes with maximum signal strength below 20 dB that would likely be missed during manual detection. These results demonstrate the ability of the synthetic data approach to train a machine learning algorithm to detect head echoes. [ABSTRACT FROM AUTHOR]