Automated Detection and Extraction of ELF/VLF Signals using Mask Regional Convolutional Neural Network

Abstract:

Extremely and very low frequency (ELF/VLF, 3 Hz-30 kHz) radio signals are generated from a variety of natural geophysical sources. In particular, signals of interest include whistler-mode chorus and hiss emissions from the space environment, and lightning generated whistlers. Although ground-based observations often contain the aforementioned signals of interest, the signals are typically immersed in a noisy environment due to lightning-generated sferics and additional man-made sources. The high noise content makes accurate extraction of relevant signals difficult, which limits the potential of large-scale statistics. Although automated detection algorithms have been employed successfully in the past, extraction of arbitrary signal classes has been a challenge. In this work, we employ a novel mask regional convolutional neural network (MRCNN) for automated extraction of arbitrary classes of ELF/VLF signals from broadband data.

MRCNN is a state-of-the-art semantic segmentation algorithm in the field of computer vision. It is a pixel-level image processing method, which is more accurate than object detection, and can automatically divide the object area from the image and identify its category. The method has been applied to VLF data by treating spectrograms as images and looking for relevant spectrographic “signals”. The program has been highly successful and hundreds of thousands of lightning-whistlers have been extracted from Palmer station data. The use of MRCNN to detect magnetospheric signals in ground data is highly robust even in the presence of noise, which is typically detrimental to more traditional time-domain automated detection techniques.

For this work, the results of lightning-whistler extraction are correlated against lightning activity in North America using the NLDN network and is shown to have a strong correspondence, further providing credence for the efficacy of the detection algorithm. Furthermore, the detected whistlers are inverted using dispersion analysis to determine magnetospheric propagation paths and cold plasma density using thousands of cases in one year. This methodology opens the door to large scale magnetospheric remote sensing without human intervention and can be easily extended to other types of signals that have established frequency-time structure.