A New Approach for the Regression of the Center Coordinates and Radius of the Solar Disk Using a Deep Convolutional Neural Network

Abstract

This paper presents a new approach for the regression of the center coordinates and radius of the solar disk in H α solar full-disk images by using a Deep Convolutional Neural Network. We use ∼100,000 original H α solar full-disk images obtained from Huairou Solar Observing Station as the experimental data set. The data set includes two parts: the original image and three numeric values (center coordinates and radius). In order to deal with the uneven distribution of the solar disk position in the original image, we randomly shift the solar disk during image preprocessing. Furthermore, data augmentation is also used to increase the robustness of the model. By evaluating the model with R -square and relative error, the center coordinates and the radius of the solar disk are proved to be effectively regressed. The data sets we constructed and source code are available as open source on GitHub.