GTC-DC 2019: GPU-Accelerated Deep Learning for Solar Feature Recognition in NASA Images

GTC-DC 2019: GPU-Accelerated Deep Learning for Solar Feature Recognition in NASA Images

Michael Kirk, NASA GSFC / CUA; James Stockton, Altamira Technologies Corporation
We’ll discuss the use of deep learning on solar images. NASA’s Solar Dynamics Observatory satellite records a 4K resolution solar image every 1.3 seconds. Scientists now have over 150 million images, or 6 PB of data, to analyze. It is challenging to identify and label classes of features common in the solar atmosphere quickly enough to deploy on the growing archive. We’ll present the first segmentation of solar images achieved through the application of a deep learning architecture. We’ll explain how we trained a convolutional neural network to segment three types of common solar features by encoding physical knowledge of the sun. We’ll describe our results and the practical knowledge gained from the application of a deep learning algorithm to a scientific problem.