Abstract

Contributed Talk - Splinter Learning

Wednesday, 23 September 2020, 09:05   (virtual room G)

Automatic Extraction of Polar Crown Filaments Using Machine Learning

A. Diercke, R. Jarolim, C. Kuckein, S. J. González Manrique, M. Ziener, A. Veronig, C. Denker
[1] AIP/Uni Potsdam, [2] Uni Graz, [3] AIP, [4] IAC/ULL/AISAS, [5] Latentine, [6] Uni Graz, [7] AIP

Polar crown filaments are prominent features on the Sun, which show a cyclic behavior aligned with the sunspot cycle. The Chromospheric Telescope (ChroTel, Tenerife, Spain) provides regular full-disk images of the Sun in three chromospheric wavelengths. One of these wavelength is H-alpha, which can be used to monitor the cyclic behavior of filaments. Manual extraction of filaments is tedious, and extraction with morphological image processing tools produces a large number of false positive detections and is slow. Automatic object detection and extraction in a reliable manner allows us to process more data in a much shorter time. We will present an overview of the ChroTel database and a proof of concept of the machine learning application, which allows us a unified extraction of, for example, filaments from the ChroTel data. As a first attempt, we approach the problem with a single shot detection algorithm from the You Look Only Once (Yolo) family of neural networks. The network is trained on the full-disk H-alpha images of ChroTel using the coordinates of bounding boxes, where the filaments are located. For ChroTel, we have 955 input images and 39782 bounding boxes. In the end, the algorithm will be standardized so that it can be applied to other full-disk H-alpha images from, e.g., the Kanzelhöhe Solar Observatory (KSO), Austria. This will be the foundation for future work developing an object detection algorithm to extract filaments from solar imaging data.