Poster Abstract

P3.8 Gaofei Zhu (The National Astronomical Observatories of the Chinese Academy of Sciences)

Theme: Data science challenges: tools from statistics to machine learning

Pixel-Level Solar Filaments Segmentation Based on Deep Learning

We presents a reliable method using deep learning to automatically segment
solar filaments in Ha full-disk solar images. This method can not only accurately identify
and segment filaments, but also minimize the effects of solar noise. First, we produce a raw
filaments dataset of thousands of images required for deep learning. Second, we propose an
automatic method for solar filaments segmentation, which is developed using U-Net based
deep convolutional networks. To test the performance of our method, a dataset composed
of 60 Ha images is considered. These images are chosen from Big Bear Solar Observatory
(BBSO) for 2013. Cross validation has shown that the proposed method can obtain promising
segmentation efficiently.