Physics-informed detection and segmentation of type II solar radio bursts - Inria - Institut national de recherche en sciences et technologies du numérique
Conference Papers Year : 2020

Physics-informed detection and segmentation of type II solar radio bursts

Abstract

Type II solar radio bursts have proven to be a useful tool for gaining insights into the behaviour of complex solar events and for forecasting and mitigating their damages on Earth. In this work, we detect and segment the occurrence of type II bursts in solar radio spectrograms, thereby facilitating the extraction of parameters needed to gain insight into solar events. We utilise prior knowledge of how type II bursts drift through frequencies over time to assist with these tasks of detection and segmentation. A new adaptive Region of Interest (ROI) is proposed, to constrain the search to regions that follow the burst curvature at a given frequency. It comes with an implicit data normalisation that reduces the variance of burst appearance in the data, hence simplifying the learning process from small datasets. We demonstrate the effectiveness of our methodology using a simple and popular HOG and logistic regression detector and basic segmentation based on voting and background subtraction. On a custom dataset representative of different levels of solar activity, at a wavelength range where no other detection algorithm currently operates, our adaptive ROI significantly improves over traditional sliding windows. In future work, it may be applied to other, state-of-the-art, machine learning algorithms.
Fichier principal
Vignette du fichier
adaptive_ROI_cameraready.pdf (1.43 Mo) Télécharger le fichier
Origin Files produced by the author(s)
Loading...

Dates and versions

hal-02923001 , version 1 (26-08-2020)

Identifiers

  • HAL Id : hal-02923001 , version 1

Cite

Joseph Jenkins, Adeline Paiement, Jean Aboudarham, Xavier Bonnin. Physics-informed detection and segmentation of type II solar radio bursts. British Machine Vision Virtual Conference, Sep 2020, Virtual, United Kingdom. ⟨hal-02923001⟩
231 View
260 Download

Share

More