Type II and Type III Solar Radio Burst Classification Using Transfer Learning

TL;DR

This paper demonstrates that transfer learning with pre-trained deep learning models can effectively classify Type II and Type III solar radio bursts from spectrogram images, aiding real-time space weather monitoring.

Contribution

It introduces a dataset of radio spectrograms and evaluates multiple pre-trained models, identifying YOLOv8 as the most accurate for automated SRB classification.

Findings

YOLOv8 achieved the highest F1 score of 92%.

Pre-trained models effectively classify SRBs with limited data.

Transfer learning offers a practical solution for space weather event detection.

Abstract

The Sun periodically emits intense bursts of radio emission known as solar radio bursts (SRBs). These bursts can disrupt radio communications and be indicative of large solar events that can disrupt technological infrastructure on Earth and in space. The risks posed by these events highlight the need for automated SRB classification, providing the potential to improve event detection and real-time monitoring. This would advance the techniques used to study space weather and related phenomena. A dataset containing images of radio spectra was created using data recorded by the Compound Astronomical Low frequency Low cost Instrument for Spectroscopy and Transportable Observatory (e-Callisto) network. This dataset comprises three categories: empty spectrograms; spectrograms containing Type II SRBs; and spectrograms containing Type III SRBs. These images were used to fine-tune several popular pre-trained deep learning models for classifying Type II and Type III SRBs. The evaluated models included VGGnet-19, MobileNet, ResNet-152, DenseNet-201, and YOLOv8. Testing the models on the test set produced F1 scores ranging from 87\% to 92\%. YOLOv8 emerged as the best-performing model among them, demonstrating that using pre-trained models for event classification can provide an automated solution for SRB classification. This approach provides a practical solution to the limited number of data samples available for Type II SRBs.