Monitoring space weather: using automated, accurate neural network based whistler segmentation for whistler inversion

It is challenging, yet important, to measure the - ever-changing - cold electron density in the plasmasphere. The cold electron density inside and outside of the plasmapause is a key parameter for radiation belt dynamics. One indirect measurement is through finding the velocity dispersion relation exhibited by lightning induced whistlers. The main difficulty of the method comes from low signal-to-noise ratios for most of the ground-based whistler components. To provide accurate electron density and L-shell measurements, whistler components need to be detectable in the noisy background, and their characteristics need to be reliably determined. For this reason precise segmentation is needed on a spectrogram image. Here we present a fully automated way to perform such an image segmentation by leveraging the power of convolutional neural networks, a state-of-the-art method for computer vision tasks. Testing the proposed method against a manually, and semi-manually segmented whistler dataset achieved <10% relative electron density prediction error for 80% of the segmented whistler traces, while for the L-shell, the relative error is <5% for 90% of the cases. By segmenting more than 1 million additional real whistler traces from Rothera station Antarctica, logged over 9 years, seasonal changes in the average electron density were found. The variations match previously published findings, and confirm the capabilities of the image segmentation technique.