Deep Learning Algorithms Applied to the Classification of Video Meteor Detections

Abstract

Abstract The application of a class of advanced machine learning techniques, namely deep learning, has been applied to automating the confirmation/classification of potential meteor tracks in video imagery. Deep learning is shown to perform remarkably well, even surpassing human performance, and will likely supplant the need for human visual inspection and review of collected meteor imagery. When applied to time series measurements of meteor track centroid positions and integrated intensities obtained from each video frame, a recurrent neural network (RNN) has achieved 98.1 per cent recall, which is defined as the number of true meteors properly classified as meteors. The RNN allowed only 2.1 per cent leakage, defined herein as the number of false positives that were incorrectly identified as meteors. The desire is to maximize recall to avoid missed orbit estimations, while also minimizing false alarms leaking through to the next processing stage of multi-site trajectory and orbit estimation. When two-dimensional spatial imagery is available or the temporal image sequence can be reconstructed, these results climb to 99.94 per cent recall and only 0.4 per cent leakage when employing a convolutional neural network (CNN). This has been further generalized from a baseline of interleaved analog video to modern progressive scan digital imagery with equivalent results. The trained CNN, nicknamed MeteorNet, will be used for post-detection automated screening of potential meteor tracks and explored in the future as a potential upstream meteor detector.