Session 7 - Machine Learning and Forecasting

Conveners: Véronique Delouille (ROB), Meng Jin (LMSAL)
Thursday 1/11, 09:00 - 10:30

Since its commissioning in 2010, SDO has captured one high resolution image every second. A large community effort resulted in the implementation of a series of algorithms for the autonomous detection and tracking of solar events, with outputs available within the Heliophysical Event Knowledgebase (HEK). Machine learning and deep learning have the ability to further leverage on these data, algorithms and database, and to tackle open complex problems. This session will discuss mathematical, statistical, computational, and machine learning techniques for information processing of SDO and solar data, with emphasis on: autonomous prediction of solar events from heterogenous data (magnetogram, EUV imagers, spectrometer), supervised and unsupervised detection, tracking of solar events and features at large and small scales, dimensionality reduction, image enhancement, methods for enriching and mining the HEK e.g. in view of spatio-temporal pattern discovery. Of special interest are techniques that allow for a progress towards reconciling theory with observations.

Plenary speaker: Vincent Barra (Université Clermont-Ferrand)

Talks

Click here to toggle abstract display in the schedule

Thursday November 1, 09:00 - 10:30

09:00	Deep Learning and solar data processing - an introduction with applications	Barra, V	Invited Oral
	Vincent BARRA
	Clermont Auvergne University, LIMOS UMR 6158 CNRS
	Over the past few years, deep learning has become the state-of-the-art technique for many problems: classification, visual recognition, speech recognition, natural language processing or even playing games. The developement of new algorithms as well as the improvment of hardware and computing capacity make these methods able to address new application fields with big data and complex problems. In this talk, I will introduce the main topics that underlie the deep learning paradigm, and then detail four possible applications that solar data processing can benefit from: classification of images or events for segmentation or semantic recognition purposes ; inverse or ill-posed problem solving, particularly focusing on image super resolution ; and data generation using generative adversarial networks.
09:30	Sunspot Group Classification using Neural Networks	Maloney, S	Oral
	S. A. Maloney, P. T. Gallagher
	Trinity College Dublin
	Sunspots are the sources of the most extreme and potentially adverse solar events such as flares and CMEs. As such many forecasting systems have been developed to predict these events, a number of which rely on sunspot group classifications. The classifications are manually produced so are subject to human errors and biases. Additionally, as the classifications are only produced on a daily basis this limits the time resolution of some forecasting methods. Further with the imaging cadence of SDO HMI, it would be impossible for a human to produce classifications for every observation. As such the development of an automated classification system would provide many benefits. Neural networks (NNs) have proven to be powerful tools for solving many complex problems such as classification, regression, and optimisation. In particular, the application of convolutional neural networks (CNNs) to image classification has greatly improved the performance of such systems. The first example of this, in the 1990s, was the identification of handwritten digits from 646 checks an 82% accuracy was achieved. Since then there have been numerous advances in both the network architectures and the underlying components. Recently an accuracy rate of 97.75% was achieved, identifying 1000 classes in 150,000 images for the ILSVRC2017 challenge. We applied a number of modern CNN architectures to the problem of classifying sunspots groups in SDO HMI observations. The input data consisted of SDO HMI SHARPs magnetograms and the daily McIntosh or Mount Wilson sunspot classifications provided by the USAF/NOAA. The entire dataset (2011-2018) was randomly split into three sets, train, test and validate. The train and test sets were used to optimise the parameters and hyperparameters of the chosen network architectures to achieve optimal performance. Once the all the parameters were fixed the accuracy of the networks were determined using the validation set containing only unseen data. We present the results of this work together with some issues encountered and avenues of further research.
09:45	AI-generated magnetogram and EUV image of the Carrington event and the estimation of its Dst value	Lee, H	Oral
	Harim Lee[1], Y.-J. Moon[1], Daye Lim[1], Eunsu Park[1], and Taeyoung Kim[1,2]
	[1]School of Space Research, Kyung Hee University, Yongin-si 17104, Korea, [2]School of Space Research, Kyung Hee University, Yongin-si 17104, Korea
	We apply an image-to-image translation model, which is a popular deep learning method based on conditional Generative Adversarial Networks (cGANs), to the generation from sunspot drawings to the corresponding magnetograms and EUV images. For this, we train the model using pairs of sunspot drawing from Mount Wilson Observatory (MWO) and their corresponding SDO/HMI magnetogram and SDO/AIA images from 2012 to 2013. We test the model by comparing pairs of actual magnetogram (EUV image) and the corresponding AI-generated one in 2014. Our results show that bipolar structures of the AI-generated magnetograms are consistent with those of the original ones and their unsigned magnetic fluxes (or intensities) are well correlated with those of the original ones with good correlation coefficients. Using this model with the Carrington sunspot drawing, we successfully produce AI-generated magnetogram (EUV image) and estimate unsigned magnetic fluxes. Using several empirical relationships (magnetic flux vs. CME speed, CME speed vs. ICME speed, and ICME speed vs. Dst) in 23 and 24th solar cycle, we conjecture the Dst value of the Carrington event, about -1,670 nT, which is similar to that of Tsurutani et al. (2003).
10:00	Solar EUV Spectral Irradiance by Deep Learning	Wright, P	Oral
	Paul Wright[1], Richard Galvez[2], Alexandre Szenicer[3], Rajat Thomas[4], Meng Jin[5,6], David Fouhey[7], Mark Cheung[6,8], Andres Munoz-Jaramillo[9], Graham Mackintosh[10]
	[1]University of Glasgow, Glasgow, United Kingdom, [2]New York University, New York, NY, United States, [3]University of Oxford, Oxford, United Kingdom, [4]University of Amsterdam, Amsterdam, Netherlands, [5]SETI Institute, Mountain View, CA, United States, [6]Lockheed Martin Solar and Astrophysics Laboratory, Palo Alto, CA, United States, [7]University of California, Berkeley, Berkeley, United States, [8]Stanford University, Stanford, CA, United States, [9]Southwest Research Institute Boulder, Boulder, CO, United States, [10]NASA Frontier Development Lab, Mountain View, CA, United States
	Extreme UV (EUV) radiation from the Sun’s transition region and corona is an important driver for the energy balance of the Earth’s thermosphere and ionosphere. To characterise and monitor solar forcing on this system and associated space weather impacts, the EUV Variability Experiment (EVE) instrument onboard NASA’s Solar Dynamics Observatory (SDO) was designed to measure solar spectral irradiance (SSI) in the 0.1 to 105 nm wavelength range. As the result of an electrical short, the MEGS-A component of EVE stopped delivering SSI data in the 5 - 35 nm wavelength range in May 2014. We demonstrate how a Residual Neural Network (ResNet) augmented with a Multi-Layer Perceptron (MLP) can fill this gap using narrowband UV and EUV images from the Atmospheric Imaging Assembly (AIA) on SDO. As a performance benchmark, we also show how our deep learning approach outperforms a physics model based on differential emission measure inversions. This work was performed at NASA’s Frontier Development Lab, a public-private initiative to apply AI techniques to accelerate space science discovery and exploration.
10:15	Statistically Identifying Systematics from Far-side Acoustic Images	Hess webber, S	Oral
	Shea A. Hess Webber[1], Junwei Zhao[1]
	[1]Stanford University
	The prediction of solar wind conditions is a critical aspect of space weather forecasting. Solar wind models are highly dependent on the global magnetic field at the solar surface as their inner boundary condition, and the lack of true knowledge of the global field in traditional synoptic maps is a significant problem plaguing space weather forecasting. Currently, only near-side magnetic field observations exist, but far-side magnetic field can be essential for accurate modeling of the Sun’s coronal field and solar wind. Current methods and observations exist that estimate the far-side active region configuration (e.g., helioseismic imaging, flux transport models, far-side EUV observations); however, each of these methods has drawbacks. Our work combines existing techniques with machine-learning and statistical analysis to develop reliable, calibrated far-side magnetic-flux maps in near-real-time, using helioseismic far-side images that are solely dependent on near-side observations. We present an overview of this method and discuss an important progress milestone: We have established a statistical relation between the far-side acoustic images from 7 years of SDO/HMI data and STEREO/EUVI 304 \AA\ data, allowing us to remove some systematics and improve the quality of the far-side acoustic images.

Posters

Click here to toggle abstract display in the list

e-Posters

Estimating solar far-side magnetic flux from STEREO EUV observations by deep learning	Chen, R
Ruizhu Chen[1,2], Junwei Zhao[2]
[1]Physics Department, Stanford University, [2] Hansen Experimental Physics Laboratory, Stanford University
The solar far-side magnetic flux is a crucial observable for space weather forecasting and solar wind modeling. However, such magnetic flux is not directly measured, but can be estimated using the far-side STEREO observations of extreme ultraviolet (EUV) 304 Å flux through the relation between the two quantities established by a machine-learning method. The near-side magnetic flux and EUV 304 Å flux are observed continuously for 8 years by the SDO/HMI and SDO/AIA, respectively, therefore, an image-to-image relation can be built by a deep-learning algorithm using these data. A deep neural network is thus trained using 10,000 calibrated image pairs of the magnetic flux and the EUV flux, sampled from the 8 years of SDO observations. The trained neural network can reproduce the magnetic flux in great details from the EUV flux and with a similarity close to 0.9. We then apply this neural network on the STEREO EUV images, and derive the maps of magnetic flux on the far side.
A deep learning approach to forecast tomorrow's solar wind parameters.	Shneider, C
Carl Shneider[1], Mandar Chandorkar[1], Monica Bobra[2], Enrico Camporeale[1]
[1] Centrum Wiskunde & Informatica, [2] Stanford University
Deep learning has proven extremely successful both in classification and regression problems, especially when it is trained on very large datasets. In the Space Weather context, despite the unarguably large amount of data at our disposal, it remains an open question whether historical datasets contain enough information to build a predictive deep learning system. In this work, we use multi-wavelength solar images from both SDO (Solar Dynamics Observatory) and SOHO (Solar and Heliospheric Observatory) as inputs to a convolutional neural network, to predict solar wind parameters observed at L1, 24 hours ahead.

p-Posters

Supervised Neural Networks for Helioseismic Ring-diagram Inversions	Hanson, C
Rasha Alshehhi[1], Chris S. Hanson[2], Laurent Gizon[2,3,1]
[1]Center for Space Science NYUAD Institute New York University Abu Dhabi UAE, [2]Max Planck Institute for Solar system research, G{\"o}ttingen, Germany, [3] Institut f{\"u}r Astrophysik, Georg-August-Universit{\"a}T G{\"o}ttingen, Germany
The inversion of ring fit parameters to obtain subsurface flow maps in ring-diagram analysis for SDO observations is computationally expensive. We apply machine learning techniques to the inversion step of the pipeline, to replace future inversion requirements. We utilize Artificial Neural Networks as a supervised learning method for predicting the flows in $15^\circ$ ring tiles. To demonstrate that the machine learning results still contain the subtle signatures key to local helioseismic studies, we use the machine learning results to re-detect equatorial Rossby waves. We find the Artificial Neural Network is computationally efficient, can achieve a root mean-square error of half that reported for the observations, and reduce computational burden by two orders of magnitude. We find that the signatures of the Rossby waves are still in the machine learning results, showing that important helioseismic signatures are maintained.
Extracting Solar Physics from 151 Million EUV Images	Kirk, M
Michael S. Kirk[1], Barbara Thompson[2], Raphael Attie[3], Nicki Viall-Kepko[2], Peter Young[4]
[1]NASA GSFC / Catholic University of America, [2]NASA GSFC, [3]NASA GSFC / NPP, [4]NASA GSFC / GMU
Beginning in 2010, the Solar Dynamics Observatory's Atmospheric Imaging Assembly (SDO AIA) revolutionized solar imaging with its high temporal and spatial resolution and coverage. The archive of extreme ultraviolet (EUV) images is now over 150 million and continues to grow. Automated algorithms consistently clean these images to remove magnetospheric particle impacts on the CCD cameras, but it has been found that compact, intense solar brightenings are often removed as well. There are now over 3 trillion "spiked pixels" that have been removed from EUV images. We estimate that 0.001% of those are of solar origin and removed by mistake – an unexplored dataset of about 30 million events. We take a novel approach and survey the entire set of AIA "spike" data to identify and group compact brightenings across the entire SDO mission. We then use the spike database to form statistics on compact solar brightenings without having to process large volumes of full-disk AIA data. The qualities of the “spikes” with a solar origin represent the most complete archive of compact EUV bright points ever assembled.
Irradiance Coronal Dimming and its Connection to CME Kinetics	Mason, J
James Paul Mason[1], Nick Arge[1], Larisza Krista[2], Alysha Reinard[3], Barbara J. Thompson[1], David F. Webb[4], Jake Wilson[5], Thomas N. Woods[6]
[1]NASA Goddard Space Flight Center, [2]University of Colorado / CIRES, Boulder, CO, USA, [3]NOAA/ SWPC, Boulder, CO, USA, [4]Boston College Institute for Scientific Research 5University of Maryland, College Park, USA, [6]University of Colorado at Boulder Laboratory for Atmospheric and Space Research
When coronal mass ejections (CMEs) depart the corona, they leave behind a transient void. Such a region evacuated of plasma is known as a coronal dimming and it contains information about the kinetics of the CME that produced it. The dimming can be so great in the extreme ultraviolet (EUV) that it reduces the overall energy output of the sun in particular emission lines, i.e., dimming is observable in spectral irradiance. We use the Solar Dynamics Observatory (SDO) EUV Variability Experiment (EVE) data to search for and parameterize dimming. We focus our search on the 39 extracted emission lines data product. We are searching these light curves for dimming around all of the >8,500 ≥C1 solar flares in the SDO era. Our method of combining these 39 light curves to remove the flare peak results in 1,521 light curves for every solar flare. Thus, we come to a total of ~13 million light curves in which to search for dimming. The question is: which ones are sensitive to CME-induced dimming? To answer this and related questions, I’m using machine learning techniques built into python’s scikit-learn library. I will describe the results of applying these techniques to the EVE data to produce the catalog, to the catalog itself, and to comparisons with other related catalogs. All of the code is open source python available on GitHub (https://github.com/jmason86/James-s-EVE-Dimming-Index-JEDI).
Coronal holes detection using supervised classification	Delouille, V
Veronique Delouille (STCE/Royal Observatory of Belgium) Veronique Delouille[1], Stefan Hofmeister[2], Martin Reiss[3], Benjamin Mampaey[1], Manuela Temmer[2], Astrid Veronig [2]]
[1]Royal Observatory of Belgium, Belgium, [2]University of Graz, Austria, [3]Space Research Institute, Graz, Austria
We demonstrate the use of machine learning algorithms in combination with segmentation techniques in order to distinguish coronal holes and filaments in solar EUV images. We used the Spatial Possibilistic Clustering Algorithm (SPoCA) to prepare data sets of manually labeled coronal hole and filament channel regions present on the Sun during the time range 2010-2016. By mapping the extracted regions from EUV observations onto HMI line-of-sight magnetograms we also include their magnetic characteristics. We computed average latitude, area, shape measures from the segmented binary maps as well as first order, and second order texture statistics from the segmented regions in the EUV images and magnetograms. These attributes were used for data mining investigations to identify the most performant rule to differentiate between coronal holes and filament channels, taking into account the imbalance in our dataset which contains one filament channel for 15 coronal holes. We tested classifiers such as Support Vector Machine, Linear Support Vector Machine, Decision Tree, k-Nearest Neighbors, as well as ensemble classifier based on Decision Trees. Best performance in terms of True Skill Statistics are obtained with cost-sensitive learning, Support Vector Machine classifiers, and when HMI attributes are included in the dataset.
Driving Scientific Discovery with Machine Learning and AI at the NASA GSFC Center for HelioAnalytics	Thompson, B
Barbara J. Thompson[1], Michael S. Kirk[1,2], Menelaos Sarantos[1]
[1]NASA Goddard Space Flight Center, USA, [2]Catholic University of America
What is HelioAnalytics? This is a broad term meant to cover all the ways that we harness advanced statistics, informatics and computer science methods to achieve our science. Our focus is on problems that we can attack with modern methods that we cannot attack otherwise. A keener understanding of how information is derived from data, and how machine learning can be harnessed to accomplish this, will expand the discovery potential for key heliophysics research topics and missions. We report on a new program to integrate modern information science, statistics, and scientific knowledge to advance the fundamental physics of connected sun-heliosphere-geospace system. The Center for HelioAnalytics is an “expert group” at NASA GSFC focusing on topics such as machine learning, neural networks, and data analytics in order to expand the discovery potential for key heliophysics research topics and missions. We define “HelioAnalytics” as a hybrid of Heliophysics + Machine Learning + Statistics + Information Design. Each of these are fields that are well developed in their own right; HelioAnalytics is the cross-disciplinary convergence of communities of physicists, statisticians, and computer scientists. HelioAnalytics is intended to foster research into advanced methodologies for heliophysical research, and to promulgate such methods into the broader community.
SHARPs and SMARPs: 22 years of solar active region data	Bobra, M
Monica Bobra
Stanford University
We study 22 years of solar active regions using measurements of the photospheric magnetic field taken by the SoHO/MDI and SDO/HMI instruments. We calculate various properties of these 13548 active regions -- such as size, flux, and topological features -- and determine statistical relationships between these properties and flaring activity. We also analyze 118 active regions from the period between May and October 2010, when both the SoHO/MDI and SDO/HMI instruments took co-temporal measurements of the photospheric magnetic field, to assess if, and how, our calculated properties vary with instrument. Our data series, called Space-weather HMI Active Region Patches, or SHARPs (Bobra et al. 2014), and Space-weather MDI Active Region Patches, or SMARPs (Bobra et al., 2018, in prep), provide seamless coverage of solar active regions spanning two solar cycles from 1996 to the present day.
Photospheric Magnetic Field Properties of Flaring vs. Flare-quiet active regions, V: Results from HMI	Leka, K
KD Leka[1,2], G Barnes [1]
[1] NorthWest Research Associates, [2] Nagoya University / Institute for Space-Earth Environmental Research
What constitutes the difference between those solar active regions that produce energetic events and those that do not? The answer no doubt lies in the state and ongoing evolution of the magnetic field. Extending this series of studies of the photospheric magnetic field as related to flare imminency, we consider daily evaluations of almost all HMI Active Region Patches (HARPS), including temporal evolution. Using the NWRA Classification Infrastructure based on NonParametric Discriminant Analysis, we evaluate not only the static characterization of the photospheric field (extending well beyond the SHARP parameters) but include coronal topology and time-series considerations, as well. Additionally, we extend the analysis beyond "global" parametrizations to describe sub-area sites which may play roles in coronal energization and event triggering. We report here on those parametrizations which best distinguish imminent flaring from imminent quiet sunspot groups.
AI-generated EUV images from SDO/HMI magnetograms by deep learning	Park, E
Eunsu Park[1], Yong-Jae Moon[1], Harim Lee[1]
[1]School of Space Research, Kyung Hee University
In astronomy and geophysics, multi-wavelength observations become very popular. Recently, several deep learning methods for image-to-image translations have been suggested and are successful for different types of transformation such as labels to street scene, black and white images to color ones, aerial to map, day to night, and sketch images to pictures. In this study, we apply an image-to-image translation model, based on conditional Generative Adversarial Networks (cGANs), to construct solar EUV images using solar magnetograms. For this, we train the model using pairs of SDO/AIA EUV image and their corresponding SDO/HMI line-of-sight magnetogram for all AIA wavelengths from 2011 to 2017 except September and October. We test the model by comparing pairs of actual SDO/AIA EUV images and corresponding AI-generated ones in September and October. We find that both real and AI-generated images are quite consistent with each other in that it is difficult for one to distinguish solar EUV images from AI-generated ones. Especially, 193 and 211 data sets have the best average correlation values (0.91) between actual EUV images and AI-generated ones for test data sets, being consistent with the idea that the origin of coronal heating is magnetic field. Using this model, we have a plan to construct solar EUV images with Kitt peak magnetograms since 1974. This methodology can be applicable to many scientific fields that use several different filter images.
Combining sparsity DEM inversions with event tracking for AIA data	Bethge, C
Christian Bethge[1], Amy Winebarger[2], Sanjiv Tiwari[3,4]
[1]Universities Space Research Association, Huntsville, AL, [2]NASA Marshall Space Flight Center, Huntsville, AL, [3]Lockheed Martin Solar and Astrophysics Laboratory, Palo Alto, CA, [4]Bay Area Environmental Research Institute, NASA Research Park, Moffett Field, CA
We apply a modified event tracking code (ASGARD - \emph{Automated Selection and Grouping of events in AIA Regional Data}) to the results from sparsity DEM inversions (Cheung et al, 2015) using AIA EUV data. Outputs are grouped regions (x/y/t) in multiple defined temperature bins that can then be correlated in space and time to track the thermal evolution of coronal structures. We show examples and an overview of the methodology.