Developing Deep Learning for Solar Feature Recognition in Satellite Images

We present results of applying a deep learning framework to the problem of segmenting important features in solar images. Since 2010, the SDO satellite has taken over 150 Million images of the sun in 9 different wavelengths of light as well as magnetic field measurements to monitor changes on the sun. A complete set of these images are recorded about once every 12 seconds. Along with this flow of data, the Heliophysics Events Knowledge-base (HEK) runs a collection of computer vision routines to detect, segment, and archive various dynamic events and quasi-static features on the sun. By combining the HEK with the stream of SDO images, we generated a labeled dataset of almost 30,000 images. Using this dataset, we are able to train, validate, and test a convolutional neural network (CNN) to segment sunspots, active regions, and coronal holes. We will present the results of this effort and estimates of its accuracy and applicability to dynamic solar events.