Gunessee, Akhil ; Marqué, Christophe ; Martínez Picar, Antonio ; Dolla, Laurent ; Delouille, Veronique

Talk presented at CESRA, University of Hertfordshire, UK on 2023-07-03

Abstract: Anyone familiar with the field of solar radio burst detection and classification knows the cumbersome task of having to go through daily observations and manually recording the type of solar radio burst occurring in the spectrograms. Past non-AI based approaches have been applied to avoid or reduce that particular hindrance but to limited success. One of those methods is through the use of an outlier detection rule such as the interquartile range (IQR) threshold method. Briefly, this method registers a burst when the brightness distribution over time of the spectrogram is higher than the sum of the third quartile with the product of 1.5 and its IQR. This simple criterion can be subject to many false positives and false negatives too, where false positives would be mainly due to RFIs. With the recent rise of big data and AI, we present a prototype model of detecting solar radio burst using YOLOv5. The initial prototype was trained on 306 images and resulted in a precision of 59.5% and a recall of 65.9%. This prototype also made us aware that this initial model iteration was susceptible to falsely detecting lightning strikes as bursts. Following these promising results and valuable lessons learnt, another around of annotations was just completed with roughly 1300 new images, and another class being added to the annotated data, that is, lightning strikes. The new annotations will allow the prospective model to not only detect if and how many solar radio bursts are present in a spectrogram but also classify which types are there. Ultimately, being able to correctly differentiate between genuine solar radio bursts and RFIs, and further classifying them into their different solar radio bursts types would be beneficial to monitoring solar activity and useful to the space weather community.

Keyword(s): Solar Radio Bursts ; Humain Radio Astronomy Station ; CALLISTO ; Interquartile Range Threshold ; Outlier Detection ; Deep Learning ; Object Detection
Links: link