# Fast Solar Image Classification Using Deep Learning and its Importance for Automation in Solar Physics

**Authors:** John A. Armstrong, Lyndsay Fletcher  
**arXiv:** [1905.13575](https://arxiv.org/abs/1905.13575)  
**Date:** 2019-05-31  
**Summary page:** https://www.summarizepaper.com/en/paper/1905.13575v1/

## Key points

- Solar physics has seen an exponential increase in data collected over the past decade
- Introduction of the Daniel K. Inouye Solar Telescope (DKIST) means we are entering an age of petabyte solar data
- Automated feature detection is crucial for post-processing solar images and creating catalogues of data ready for researchers to use
- John A. Armstrong and Lyndsay Fletcher proposed a deep learning model using a convolutional neural network to accomplish this task efficiently
- The authors trained their network using Hinode/Solar Optical Telescope (SOT) H$\alpha$ images of a small subset of solar features with different geometries such as filaments, prominences, flare ribbons, sunspots, and the quiet Sun
- The network achieved near-perfect performance on classifying unseen images from SOT (approximately 99.9%) in just 4.66 seconds
- Transfer learning was explored by using pre-trained deep neural networks to help train new models for the first time in a solar context
- The authors showed that their network is robust to changes in resolution without affecting its performance
- However, the network fails to generalize to sunspots from SDO/AIA bands 1600/1700\AA{} due to small-scale brightenings around them and prominences in SDO/AIA 304\AA{} due to coronal emission
- Automated feature detection through deep learning models like this one will be invaluable for post-processing massive amounts of solar data collected by DKIST and other solar observatories

## Summary

Solar physics has seen an exponential increase in the volume of data collected over the past decade, and with the introduction of the Daniel K. Inouye Solar Telescope (DKIST), we are entering an age of petabyte solar data. To post-process solar images and create catalogues of data ready for researchers to use, automated feature detection is crucial. A deep learning model proposed by John A. Armstrong and Lyndsay Fletcher can accomplish this task efficiently using a deep convolutional neural network that is adept at feature extraction and image processing. The authors trained their network using data from Hinode/Solar Optical Telescope (SOT) H$\alpha$ images of a small subset of solar features with different geometries such as filaments, prominences, flare ribbons, sunspots, and the quiet Sun (i.e., absence of any other four features). The network achieved near-perfect performance on classifying unseen images from SOT (approximately 99.9%) in just 4.66 seconds. For the first time in a solar context, transfer learning was explored by using pre-trained deep neural networks to help train new models. The authors showed that their network is robust to changes in resolution by degrading images from SOT resolution (approximately 0.33$^{\prime \prime}$ at $\lambda$=6563\AA{}) to Solar Dynamics Observatory/Atmospheric Imaging Assembly (SDO/AIA) resolution (approximately 1.2$^{\prime \prime}$) without affecting its performance. However, it was observed that the network fails to generalize to sunspots from SDO/AIA bands 1600/1700\AA{} due to small-scale brightenings around them and prominences in SDO/AIA 304\AA{} due to coronal emission. Automated feature detection through deep learning models like this one will be invaluable for post-processing massive amounts of solar data collected by DKIST and other solar observatories.