Computer Vision for the Solar Dynamics Observatory · Computer Vision for the Solar Dynamics Observatory ... produce a comprehensive system of automated feature recognition for the

Solar PhysicsDOI: 10.1007/•••••-•••-•••-••••-•

Computer Vision for the Solar Dynamics Observatory

P.C.H. Martens1· G. Attrill1 · A.R. Davey1

·

S. Farid1· P.C. Grigis1

· J. Kasper1·

K. Korreck1· S.H. Saar1

· Y. Su1·

A. Savcheva1· P. Testa1

· M. Wills-Davey1·

P.N. Bernasconi2 · M.K. Georgoulis2·

V.A. Delouille3· J.F.. Hochedez3

·

J.W.. Cirtain4· C.E,. DeForest5

·

R.A. Angryk6· I. De Moortel7 ·

T. Wiegelmann8·

c© Springer ••••

Abstract In the fall of 2008 NASA selected a large international consortium toproduce a comprehensive system of automated feature recognition for the SDOmission. The SDO data we consider are all AIA images plus surface magneticfield images from HMI. Helioseismology is addressed by another group. We willproduce robust and very efficient professionally coded software modules thatcan keep up with the relentless SDO data stream and detect, trace, and analyzea large number of phenomena, including: flares, sigmoids, filaments, coronaldimmings, polarity inversion lines, sunspots, X-ray bright points, active regions,coronal holes, EIT waves, CME’s, coronal oscillations, and jets. In additionwe will track the emergence and evolution of magnetic elements down to thesmallest features that are detectable, and we will also provide at least four fulldisk nonlinear force-free magnetic field extrapolations per day. The detection ofCME’s and filaments is accomplished with SoHO/LASCO resp. ground basedHα data.

1 Harvard-Smithsonian Center for Astrophysics, 60 GardenStreet, Cambridge, MA 02138, USA email:[email protected] Johns Hopkins University/Applied Physics Laboratory,11100 Johns Hopkins Road, Laurel, MD 20723, USA3 SIDC-Royal Observatory of Belgium, Avenue Circulaire 3,B-1180 Brussels, Belgium4 Marshall Space Flight Center-NASA, Mail Code:VP 62,Marshall Space Flight Center, AL 358125Southwest Research Institute, 1050 Walnut Street Suite300, Boulder, CO 80302, USA6Department of Computer Science, Montana StateUniversity, 362 EPS, Bozeman, MT 59717-3880, USA7School of Mathematics & Statistics, University of StAndrews, North Haugh, St Andrews, KY16 9SS, UK8Max-Planck-Institut fuer Sonnensystemforschung,Max-Planck Strasse 2, 37191 Katlenburg-Lindau, Germany

SOLA: computer-vision09.tex; 19 May 2009; 15:59; p. 1

Martens et al.

A completely new software element to round out this suite is a trainable fea-ture detection module that employs a generalized image classification algorithmto produce the texture features of the images analyzed. We have successfullytested a proto-type on TRACE data. Such a trainable module can be used tofind features that have not even been discovered yet (as, for example, sigmoidswere in the pre-Yohkoh era). Our codes will produce entries in the HeliophysicsEvent Knowledge Base. This will permit users to locate data on individual eventsas well as carry out statistical studies on large numbers of events, using theinterface provided by the Virtual Solar Observatory.

The operations concept for our computer vision system is that the data willbe analyzed in near real time as soon as they arrive at the SDO Joint ScienceOperations Center and have undergone basic processing. This will allow thesystem to produce timely space weather alerts, and to guide the selection andproduction of quicklook images and movies, in addition to its prime mission ofenabling solar science. A complex and unique data processing pipeline consistingof the hardware and control software required to handle the unprecedented SDOdata stream and accommodate the computer vision modules is being set up atLMSAL by the JSOC team, with an identical copy at SAO.

Keywords: Instrumentation and Data Management

1. Introduction

The open and immediate dissemination of solar data from all NASA solar mis-sions is one of the greatest assets in solar physics, and the Solar DynamicsObservatory (SDO) represents a new frontier in quantity and quality of solardata. At about two TB/day, the data will not be easily digestable by solarphysicists using the same methods that previous missions have employed. Theavailability of this data in a form that is useful for scientists will be crucial tothe success of the mission.

In order for solar scientists to use the SDO data effectively they need metadatathat will allow them to identify and retrieve data sets that address particularscience questions. Providing this metadata via the pipeline processes describedbelow, is the core of our computer vision project.

We are building a comprehensive computer vision post-processing pipelinefor SDO, abstracting complete metadata on the features and events detectableon the Sun without human intervention. Our feature finding team will debug,deploy, and support a modular, extensible pipeline framework to augment theexisting data system, and comprises experts in each subfield of computer vision.

Our computer vision project unites more than a dozen individual, existingcodes into a systematic tool that can be used by the entire solar analysis com-munity. In addition to static, standards-based codes that detect well defined andwell studied features, we will develop and deploy a trainable feature recognitionsystem for post-processing, which will enable the systematic study of human-recognizable features that have not yet been identified. This unique capabilityallows flexible scientific exploration of the SDO data set as new types of featuregain interest.


Computer Vision for SDO

2. Operations Concept and Overview of Modules

The feature finding codes described here will be part of the SDO Event DetectionSystem (EDS) at the Joint Science Operations Center (JSOC; joint betweenStanford and LMSAL). The basic purpose of the EDS is to analyze the SDOimagery for events and features as soon as the data reach the JSOC, i.e. in near-real time. This concept implies that the analysis has to be automated. The set-upallows for the generation of space weather alerts, relevant quicklook images andmovies, as well timely alerts for the Solar Physics community. The feature findingmodules that are described in the remainder of this paper form the core part ofthe EDS, but there will also be room for other, community produced tools.

The metadata produced by the software modules that are part of the EDSwill be stored in the Heliophysics Event Knowledge Base (HPKB), which will beeasily accessible on-line for the rest of the world via the Virtual Solar Observatory(VSO) . Solar scientists will be able to use the HPKB to select event and featuredata to download for science studies. Such a preselection is necessary becauseit is impossible to download and store the full 1 TB/day of SDO imagery. SDOdata will be available for downloading from archives at JSOC, SAO, and possiblya European site.

In order for our feature finding algorithms to survive the firehose of SDO datawe need to produce software that is modular – i.e. a separate program for eachfeature finding task – robust, and efficient. The latter two requirements aremet by using professional programmers: our resource intensive modules, afterinitial design by scientists, will be coded and tested by the same team at SAOthat is responsible for the Chandra software and archive.

Some of the feature finding modules that are described in the remainder of thispaper will be continuously monitoring the data stream and extracting metadata,while others (such as the flare module) are self-triggered by continuously moni-toring part of the data stream for an alert, and then springing in full action whensuch an alert is issued. A third group of modules does not operate continuously,but is triggered by one or more alerts from other modules (e.g. wave detectionalgorithms).

In Table 1 we present an overview of the software modules to be produced byour Feature Finding Team (FFT), specifying their names, modes of operation,scientific purpose, responsible Co-I’s, etc. A much more detailed individual de-scription of each module is given in the remainder of this paper. At the end wealso describe the hardware setup of the data pipeline in which the modules willbe operating. Identical copies of the data pipeline have been installed at LMSALand SAO.


Martens et al.

Table 1. Elements of the SDO Computer Vision system.

Science Target Flares Filaments Sigmoids CMEs Coronal Dimming Jets

S/W Module Name Flare Detective AAFDCC Sigmoid Sniffer CME Detector/Tracker Jet Detector

Triggered or Continuous? Triggered Continuous Triggered Triggered Triggered Triggered

Trigger Source Data AIA 193 AIA 211 LASCO C2 and C3 AIA 193 AIA 304

Trigger Source Module(s) Self Self Flares, Dimmings, Sigmoids Self XRBP & CH

Binned Image for Trigger 16 X 16 2048 x 2048 512x512 1024x1024

Cadence for Trigger Full 10 min Full LASCO Cadence 10 min 1 min

Source Data All AIA Channels Global Hi-res Ha Network AIA 94, 131, 211, 335 LASCO C2 and C3 AIA AIA 304

Source Data Binning None? Full; some smoothing during data processingNone None; front smoothing for analysis1024x1024 Full

Source Data Cadence Full Full (~ 4/day eventually) 10-20 s Full (~1 hr) 10 min 1 min

Outputs

Output Data Volume ~ 1.5 MB per Ha image

S/W Module Data Flare Detective AAFDCC Sigmoid Sniffer CME Detector/Tracker Jet Detector

Programming Language IDL IDL and C IDL IDL IDL IDL

Module Location LMSAL SAO LMSAL SAO LMSAL SAO

Computing Requirement Most < 1 min CPU per Ha Image

Originator Paolo Grigis Pietro Bernasconi Manolis Georgoulis Meredith Wills-Davey Gemma Attril Antonia Savcheva

Originating Organization SAO JH-APL JH-APL SAO SAO SAO

Priority #1 #4 #2

Inauguration 31-Aug-09 Fall 09 31-Aug-09

Heritage EIT, TRACE GBO H-alpha SXT, SXI LASCO New XRT

Science Target Oscillations EIT Waves Active Regions Coronal Holes X-ray Bright Points

S/W Module Name Osc. Finder EIT Wave Tracker SPoCA SPoCA BP Finder

Triggered or Continuous? Triggered Triggered Continuous Continuous Continuous

Trigger Source Data AIA 193 AIA 193

Trigger Source Module(s) Flares, CMEs, Dimmings Flares, CMEs, Dimmings

Binned Image for Trigger N/A

Cadence for Trigger N/A Full

Source Data Appropriate AIA Channels AIA AIA 171/195, 211/335, 94 AIA 171/195, 211/335, 94 AIA 171, 195, 211

Source Data Binning None None; smoothing applied None None None?

Source Data Cadence None during analysis Full Full Full

depends on frequency of parameter updates

Outputs

Output Data Volume

S/W Module Data Osc. Finder EIT Wave Tracker SPoCA SPoCA BP Finder

Programming Language IDL IDL C++ C++ IDL

Module Location SAO SAO LMSAL LMSAL SAO

Computing Requirement

Originator Ineke De Moortel Meredith Wills-Davey Veronique Delouille Jean-Francois Hochedez Steve Saar

Originating Organization St Andrews SAO ROB ROB SAO

Priority #5

Inauguration

Heritage TRACE, EIT EIT EIT EIT EIT

Science Target Magnetic Feature Tracking Sunspots PIL Mapping Global NLFFFs Trainable Feature Recognition

S/W Module Name SWAMIS SWAMIS PIL Finder Optimization Code for Full Disk

Triggered or Continuous? Continuous Continuous Continuous Continuous Continuous

Trigger Source Data

Trigger Source Module(s)

Binned Image for Trigger

Cadence for Trigger

Source Data HMI Time Averaged Magnetograms HMI Magnetograms HMI Magnetograms HMI+AIA for comparison All AIA Channels, some HMI Data

Source Data Binning None None Depends on science specs Binned to 256x256 for ARs 128X128 component analysis

Source Data Cadence 6 minutes TBD (5-60 min) ~5 min ~5 min Cadence TBD

Outputs 120 MB/day nominal < 1 MB/day 3D-data-cube 10-16 texture parameters per component

Output Data Volume 256x256x256

S/W Module Data SWAMIS SWAMIS PIL Finder Optimization Code Trainable Feature Recognition

Programming Language PDL PDL IRAF C C/IDL

Module Location JSOC LMSAL MPS SAO

Computing Requirement 8-16 CPUs to operate at 4x realtime < 1 CPU

Originator Craig DeForest Craig DeForest Meredith Wills-Davey Thomas Wiegelmann Rafal Angryk

Originating Organization SwRI SwRI SAO Max Planck Sonnenforschung MSU

Priority #3

Inauguration 31-Aug-09

Heritage MDI MDI Kitt Peak SOT TRACE

3. Flare Detection

The flare detection module of the computer vision system serves a twofoldpurpose:

• Provide rapid flare alerts for the space weather community in near real-time(trigger component)



• Generate a statistical survey of flares and measure physical parametersrelevant for flare science (analysis component)

Because this is similar to police work of quickly catching villains and thenexamining them in detail, our module is named flare detective (FD).

The trigger component of the FD needs to be very fast and works on heavilybinned AIA images of 16 × 16 macropixels, each covering approximately an80′′ × 80′′ square. A peak detection algorithm will be applied to the integratedsignal in each macropixel, thus providing the start, peak, end time, and ap-proximate location of the flare. This approach allows detection of simultaneousflares in different active regions. We will use a detection procedure based onthe flare identification algorithm used on RHESSI data by Christe et al. (2008)which works well for noisy and background-affected lightcurves. Basically, thisalgorithm smooths the lightcurves and detects flares as intervals of positivederivative and negative derivative around a local maximum. We plan to runthe trigger in the 193 A band, because its behaviour is well known from pastobservations. However, if during commissioning it is found that some flares arebetter observed in the new, hotter bands 131 A and 94 A, we will use them aswell.

The more computationally intensive analysis component will run only on thesubset of data close in time and space to the flare (as detected by the triggercomponent). The key parameters that will be determined for each flare are:flare timing, location, area, peak intensity, plasma temperature, and emissionmeasure, light curves in each EUV channel, characteristic time scales, associatedNOAA AR number, GOES class, peak emissivity in selected EUV and SXRemission lines from EVE, movies of the flare from rise to end, and a classificationof the flare in terms of flare type (e.g., two-ribbon flare, etc...). The classificationwill be provided by the generic feature recognition system, described in Section16 of this paper.

The analysis component will determine the flare area and centroid from dif-ference images at peak time vs. flare onset. A set of background-subtractedlightcurves in all EUV channels will be generated for the flare interval. Theselightcurves will be used to estimate the plasma temperature, and to provideapproximate emission measure values. Summary movies will be produced foreach event in the 171A, 193A, 94A, and 131A channels. Our proposed approachwill provide complete coverage at least down to GOES C1 level flares, withpossible extension to smaller events.

The set of statistical information about the flares provided by the FD overa substantial fraction of a solar cycle can be used to determine the distibutionof flares in space, time, location, size and thermal energy. This information is ofcritical importance to study the energy release mechanisms in flares.

4. Hα Filaments

Filaments in the solar chromosphere are well-known large-scale structures ofrelatively dense and cool plasma suspended in the hot and thin corona. They are


Martens et al.

Figure 1. Left : TRACE image of an active region segemented in macropixels, with a C1 flareright of the center of the image. Right : Lightcurves for each macropixel.

particularly well visible in Hα filtergrams. Filaments and their source, filamentchannels, are known to align with photospheric magnetic polarity inversion lines(PILs - Martin, 1998)). All solar eruptions occur above PILs. In addition,filaments are known to involve helical magnetic fields, twisted beyond theirminimum-energy, current-free, magnetic configuration (Martin, Bilimoria, andTracadas, 1994; Rust and Martin, 1994; Pevtsov, Balasubramaniam, and Rogers,2003). Non-potential (i.e., helical) magnetic fields are invariably involved in solareruptions and give rise to coronal mass ejections (CMEs). Filaments themselvesoften erupt fully or partially into CMEs, leading to a complete or, respectively,partial filament disappearance from the solar disk (Gilbert et al., 2000; Asaiet al., 2003; Jing et al., 2004). If one knows the sense of twist (chirality) ina filament before its disappearance, then one has additional clues about themagnetic helicity of the CME that might be useful in assessing the CME’spossible geoeffectiveness (e.g. Yurchyshyn et al., 2001; Rust et al., 2005).

Filaments are easily discernable by visual inspection of Hα images but non-trivial to infer, track and detect their disappearence automatically. The “Ad-vanced Automated Filament Detection and Characterization Code” (AAFDCC)that we have develped takes a step beyond a typical filament identification code.Besides mere identification, (i) it determines the filament’s complete shape, spineand orientation angle, (ii) if a filament is broken up in to two or more pieces itcorrectly identifies them as a single entity, (iii) it finds the filament’s magneticchirality (sense of twist), and (iv) tracks filaments from image to image for aslong as they are visible in the solar disk. The code was originally developedby Bernasconi, Rust, and Hakim (2005) and is fully functional, tested, andvalidated. It currently runs on an APL server on a daily basis and so far hasfully automatically generated a database of daily solar filament properties fromJuly 2000 until present. With only occasional gaps due to lack of Hα images forspecific days. To our best knowledge, our code is the only one that accomplishesall these important tasks at once without any human intervention.



We have described the algorythm extensively in Bernasconi, Rust, and Hakim(2005), so here we will only highlight its most important steps. (1) The codedownloads new full-disk Hα images from the Global High-Resolution Hα net-work (http://www.bbso.hjit.edu/Research/Halpha). (2) It determines filament-like dimmings in by using a sophisticated algorithm that combines thresholding,and morphological filtering, and detects and eliminates round features that mostlikely are sunspots, not filaments. (3) For each identified dark structure (filament)it determines the chain code outlining the shape, and identifies its “spine” bymeans of a sophisticated algorithm described in detail in (Bernasconi, Rust, andHakim, 2005). The algorithm is even capable of connecting different filament“segments” together thus uncovering the true shape of the spine. (4) Once a spineis identified, the code checks for filament barbs whose orientation (left/right-handed) implies the magnetic chirality of the filament (dextral/sinistral, respec-tively - Martin (1998)). The helicity of a given filament is determined if a decisivemajority of barbs have a common orientation and is left undetermined if thenumber of right bearing and left bearing barbs is evenly split. (5) Finally thecode compares the location of the filaments found on the most recent image withthe locations of filaments detected in previous images. This allows (i) to trackthe evolution of a filament in time; (ii) to detect the appearence of a new filament(in this case a new unique filament ID is given); (iii) to detect if a filament hasdisappeared from the visible disk thus indicating a possible filament eruption. Atypical result is shown in Figure 2. The algorithm is fast enough to operate inreal-time on current standard desktop computers, and benefits from observationswith good cadence to identify filament disappearances.

We have validated our code by comparing its results with the filament list ofPevtsov, Balasubramaniam, and Rogers (2003). The list was reproduced with anaccuracy of 72%. The main difference between our results and those of Pevtsovet al. was that a larger number of filaments had undetermined chirality in ourcase. Being able to identify the orientation of all filament barbs in all cases,we attribute this discrepancy to our unbiased automated chirality finder, asopposed to the biased chirality determination performed by a human operator(see Bernasconi, Rust, and Hakim, 2005).

For the SDO computer vision project we will modify the code to furtherimprove its filament identification and characterization performance. In partic-ular we will apply an adaptive threshold method that changes the thresholdfor identification of the filament’s outline. This will allow us to identify andcharacterize thinner filaments in active regions, that usually tend to be difficultto detect by using a single threshold for the entire solar disk. Once the basecode is operative in the pipeline we will investigate means to adapt it to detectfilaments from AIA imagery. We will also investigate a modification to the codeto identify prominences beyond the disk edge from the AIA 304 A channel.

The code output will be: (1) Full-disk Hα maps with contours of the identifiedfilaments, location of filament center, filaments spine and barbs, and printedfilament ID and its chiralily (if any determined). (2) For each filament the codewill produce a VOEvent metadata entry which will contain information about thefilament location, outline (chain code), spine skeleton, area and length, number ofbarbs and in which direction they point, and finally the filament chirality, based


Martens et al.

Figure 2. Typical results of the filament finder algorithm of Bernasconi et al., (2005), appliedto an Hα image taken by the Big Bear Solar Observatory’s (BBSO) Hα telescope at 15:36UT on 2000 November 16. Insets at the sides show three well-formed filaments magnified -one sinistral (red box) and two dextral (magenta and cyan boxes). Yellow curves outline thefilaments, cyan curves indicate their spines, and the red points indicate the intensity-weightedcentroid of the filaments.

upon the barbs direction analysis. The Filament Detection code will generateapproximately 6 MB/day of images and metadata in the VOEvent format whenrun at four images per day. Each individual run will take approximately twominutes on a medium speed workstation.

Since the SDO data do not include full-disk Hα images, we will apply ourcode to nearly simultaneous Hα images provided by the Global High-ResolutionHα network. The chain code outlines of the filaments, as well as their spine andother filament paramenters can afterwards be overplotted on any given full-diskSDO image, be it a HMI magnetogram or an AIA EUV image.

The filament code can be used for statistical studies of filament propertiesover long periods of time spanning a full solar cycle, and if run in near real-time it can provide useful information for space-weather forecasting. When thecode detects the disappearence of a large filament it can deliver a CME warningtrigger. Where the code has determined the chirality of the disappeared filamentit can give information about the orientation of the erupting flux-rope and thusthe geoeffectiveness of the associated CME.



5. Sigmoids

The solar X-ray corona often presents us with bright sigmoidal structures. Thesesigmoids are considered directly or indirectly related to magnetic structurestwisted beyond their lower-energy state (see Gibson et al., 2006; Green et al.,2007, for reviews) and, as such, they are broadly recognized as tell-tale sig-natures of unstable coronal magnetic flux systems. This view is supported bythe realization that X-ray sigmoids are associated to solar eruptions (Rust andKumar, 1996) with the great majority of them (∼ 84%, according to Canfield etal.,1999), triggering solar eruptions within one or two days. Eruptions typicallyresult in the disappearance of a sigmoid and the launch of a CME that is thoughtto be associated to the twisted/stressed sigmoidal pre-eruption fields.

Not all CMEs are preceded by sigmoids. Further, not all sigmoids are thesame. The most persistent and fainter of them may or may not trigger aneruption while the brightest, transient sigmoids show a much more robust as-sociation with eruptions (Gibson et al., 2006). From this viewpoint, althoughonly a fraction of CMEs is associated with sigmoids, it has been accepted thatan efficient, automatic sigmoid recognition offers an unbiased way of identifyingshort-term precursors for many, if not most, CMEs.

What is perhaps more important than CME prediction is the physical un-derstanding of what is a X-ray sigmoid and why it occurs. The classical view(Rust and Kumar 1996) treats sigmoids as helical flux ropes undergoing thehelical kink instability. More recent views, assisted by numerical simulations,consider upward or downward kinking, inverse bald-patch magnetic geometries,or hyperbolic flux tubes to infer sigmoids with helicity similar or opposite tothat of a Titov and Demoulin (1999) flux rope undergoing either the helical kinkinstability or the torus instability (e.g., Fan and Gibson 2003; 2004; Kliem etal.,2004; Torok and Kliem, 2005; 2007; Gibson et al.,2006; Green et al.,2007).Regardless of their complicated similarities and differences, all above models tendto treat sigmoids as single flux ropes or current sheets surrounding these fluxropes. This is being challenged by recent observations, however: of the 107 X-raysigmoids that Canfield et al.,(2007) studied using full-resolution observationsby the Soft X-ray Telescope (SXT; Tsuneta et al.,1992) onboard the Yohkohmission, none appeared to be a single S- or inverse S-shaped loop. Instead,they appeared to consist of multiple loop patterns arranged such that they arediscerned as single sigmoidal structures in limited-resolution images. If theseresults are confirmed, many sigmoid models may need substantial revision.

To the purposes of understanding sigmoids and predicting CMEs we will usethe automatic pattern recognition “sigmoid sniffer” algorithm, first describedby LaBonte et al.,(2003). A detailed description of the algorithm is provided ina companion paper in this volume (Bernasconi and Georgoulis 2009). Full-diskX-ray images are inserted as input to the algorithm that first uses multiplebrightness thresholds to discern persistent bright structures. If one or morecandidate sigmoids are identified, the code infers the orientation angle profile ofsuccessive points along the structures’ outline and compares it to the expectedprofile of the theoretical S-shaped curve. From the fit one recovers the handed-ness (forward S [right-handed], or inverse S [left-handed]), the orientation, and


Martens et al.

(a)(b)

(c) (d) (e)

Yohkoh

Figure 3. An example result of the sigmoid sniffer: (a) a right-handed sigmoid is identifiedin a Yohkoh/SXT full-disk X-ray image. (b) The location of the sigmoid (corresponding toNOAA AR 9704) and the CME warning (encircled). The sigmoid disappeared less than 30minutes later in an eruption and a subsequent CME - encircled in (c-e) - was detected bySOHO/LASCO.

the aspect ratio of the sigmoid. The code further issues a CME warning in caseit identifies a sigmoid. This information, along with the location and the totalsize of the sigmoid will be provided as output of the module.

An example output is depicted in Figure 3a. The algorithm detected a sigmoidin NOAA active region (AR) 9704 using a full-disk Yohkoh SXT image taken at14:28 UT on 2001 November 21. A CME warning was issued as a result (Figure3b). The sigmoid was transient and rapidly evolving, so a CME originating fromthis area and resulting in the disappearance of the sigmoid was detected ∼ 30minutes later, at 14:54 UT (Figures 3c-e).

The sigmoid sniffer algorithm has been applied extensively to full-disk X-rayimages from Yohkoh/SXT and the Solar X-Ray Imager (SXI) GOES mission,resulting in the automatic identification of numerous sigmoids. For this project,it will be applied to data from the following instruments:

(1) Hinode’s X-Ray Telescope (XRT; Golub et al.,2007). XRT images have sig-nificantly higher spatial resolution (1 arcsec per pixel) than Yohkoh/SXTimages so the sigmoid sniffer should be able to detect more crisp, finer sig-moids in Hinode/XRT images. Figure 4 shows a comparison between typicalHinode/XRT and Yohkoh/SXT images. The XRT CCD contamination (E.DeLuca, SolarNews, 2007 September 4 issue) covering < 4% of the XRT field



Figure 4. Comparison between a Hinode/XRT and a Yohkoh/SXT image taken at verysimilar solar activity levels of solar cycles 22 and 23 (JAXA news release of 2006 October 31).Notice the remarkable additional activity detected in the XRT image.

of view should not be a major concern for sigmoid detection, provided that thecontaminated pixels do not accumulate to form sizeable clusters. The recentlyreported irregularities in the Hinode downlink, however, may somewhat affectthe cadence of the incoming HXT images.

(2) SDO’s Atmospheric Imaging Assembly (AIA). Here we will identify sigmoidsin both the higher energy channels (94 A and 131 A), that correspond toflaring active regions (T ∼ 106.8−7.2 K), and the lower energy channels (211A and 335 A), that correspond to the (not necessarily flaring) active-regioncorona (T ∼ 106.3−6.4 K)

The high-resolution Hinode/XRT and SDO/AIA observations may help re-solve the controversy on the morphology of sigmoids as single, monolithic, ormultiple structures. In addition, it will be very interesting to see whether thesigmoidal morphology persists in different AIA temperature channels. Seriousmorphological differences in different temperatures would again constrain themodeling of sigmoids, especially the spatial distribution of temperature enhance-ments along them. For example, is the transient sigmoid caused by upwardkinking (peak temperature on top), giving rise to a classical flare cusp (Masudaet al.,1994) or is it mostly due to a downward kinking toward a possible low-lying bald patch (peak temperature on bottom)? Moreover, what is the physicaldifference between a complete and a partial sigmoid eruption, where the sigmoidreappears shortly after its disappearance?

The sigmoid sniffer will be cross-correlated with the filament identificationand characterization module of Section 4. In an illustrative example, Rust andLaBonte (2005) applied first the sigmoid sniffer to a Yohkoh/SXT image toidentify a sigmoid and then applied the filament code to it to identify its “spine”and infer its aspect ratio more accurately. The example has been reproduced inFigure 5.


Martens et al.

Figure 5. Example of the interaction between the sigmoid sniffer and the filament identi-fication and characterization module. The background is a false-color Yohkoh/SXT image,with the sigmoid appearing dark. The white box is drawn by the sigmoid sniffer while thecurved sigmoid outline and the pink sigmoid “spine” stem from the filament identificationcode (adapted from Rust and LaBonte, 2005).

The sigmoid sniffer code will generate approximately 5 MB of metadata perday in the VOE format. For each run, it uses approximately ten minutes of CPUtime on an average desktop workstation.

6. CME Recognition and Tracking

CME recognition and tracking methods, such as the SOHO CDAW database(Yashiro et al. 2004), CACTus (Robbrecht and Berghmans 2005), and the work ofOlmedo et al. (2008), have been available for some time. Our algorithm addressesthe tracking of CMEs differently than currently available methods, in that itrecords CME front dynamics organically in two dimensions. It then becomespossible to measure velocities and accelerations in the lateral as well as theradial dimensions, providing us with the ability to measure the total expansionof the CME; this is particularly useful at lower altitudes, when the base of theCME is still expanding.

CMEs are detected as they emerge from behind the edge of the LASCO C2 andC3 coronagraph disks. We find them by examining the edge of the coronagraphdisk for sudden high threshhold intensity increases from frame to frame. Thefront edge of the CME is measured in polar coordinates at the strong gradientbetween the CME brightness and the background.

Once the initial intensity increase is detected, potential lateral expansion istaken into account by continuing to look for intensity increases along adjacentradial coordinates. The entire leading front is then tracked in polar coordinates asthe event travels across the field-of-view. Unlike other tracking codes, this algo-rithm records the two-dimensional position of the leading CME front, regardlessof how the shape of the CME changes over time.

Having measured the CME front as a function of time, it becomes possibleto find two-dimensional velocity trajectories via Huygens plotting. A smoothingkernel is used to create smooth velocity curves, and the resulting data are dividedinto constant time intervals, allowing us to track any one point smoothly forward



in time. From this output, we can calculate plane-of-sky velocity and accelerationat any point along the CME front.

WIth the CME measured dynamically in two dimensions, it also becomespossible to find an average angular width, both for the entire event and as afunction of time. It is not unusual, particularly for off-center halo CMEs (Sheeleyet al. 1999), for the brightest component of the CME to not bisect the angularwidth. Therefore, to get a better sense of the true direction of propagation, wecalculate the first moment of the CME as a whole to find its “center of intensity”position as a function of time.

In cases where the CME is isolated, meets a sufficient brightness threshold,and has an obvious near-side source region (such as a flare or a dimming region),one can use the central axis of the associated “ice cream cone” model (Fisherand Munro 1984; Leblanc et al. 2001, Michalek and Gopalswamy 2003; Xie etal. 2004; Xue et al. 2005) to map its three-dimensional direction of propagation.By assuming that the CME has cylindrical symmetry, it becomes possible toproject the CME in three dimensions. Using this information, we can calculate athree-dimensional velocity vector, a quantity far more valuable than the oft-usedplane-of-sky CME vectors. This three-dimensional velocity, when combined withCME mass, will allow for the calculation of CME kinetic energies.

Determining the CME mass requires a knowledge of the volume and density ofthe event. By fitting the “ice cream cone” model to the angular width and leadingCME front, one can calculate a rough volume for the CME. Since the plasmais optically thin and any LASCO intensity signal is linear in electron columndensity, the observed Thomson-scattered intensity of a CME is proportional toits column density brightness. Using a base difference imaging technique, wecan find the excess heliospheric density due to the CME to within a factorof two (DeForest et al. 2001). We note that Webb (2000) and Bemporad et al.(2007) have demonstrated that the mass of a CME is not constant, but increasesover time. Therefore, mass (and possibly kinetic energy) will be calculated as adynamic quantity.

The algorithm we use for extracting these metadata is based on a well-established tracking and measurement code, written in IDL (Wills-Davey 2006).These methods are run on calibrated, background-subtracted SoHO-LASCOcoronagraph data. The tracking module monitors the data steam continuously,in order to catch the appearance of a CME from behind the coronagraph disk. Incases of isolated, well-defined CMEs, metadata outputs are extracted. The meta-data parameters will be converted to VOEvent format (e.g. CME AngularWidthand CME Mass) for dissemination via the HPKB. We expect to observe between0.5 and four CMEs/day, and each event will take ∼ 30 minutes to run (after theCME has been observed in its entirety). This workload can be run on a standarddesktop computer, and we expect disk space usage on ∼ 1 − 8 Mb/day.

7. Coronal Dimming Regions

The coronal dimming detection and metadata extraction algorithm together withthe flare detection and flux emergence algorithms are designed to provide space


Martens et al.

weather alerts in near real real time. The relationship between coronal dimmingsand coronal mass ejections (CMEs), acknowledged since the 1970s (e.g. Rust andHildner, 1976), has recently been confirmed statistically Bewsher, Harrison, andBrown (2008). Coronal dimmings have been noted as a reliable indicator of front-side (halo) CMEs, which can be difficult to detect in white-light coronagraphdata. In the absence of a coronagraph (such as with SDO), coronal dimmings canact as an important indicator of the launch of a CME. The algorithm describedin detail in Attrill and Wills-Davey (2009) (companion paper, this issue), isdesigned to detect and extract coronal dimming signatures in the low corona,which are associated with eruptive events, such as CMEs.

This algorithm has two main components: i) detection and ii) metadata ex-traction. The detection relies on the analysis of statistical properties of EUVimages in which coronal dimmings occur. This method was first introducedby Podladchikova and Berghmans (2005). Our implementation differs in thatit is adapted to run on non-differenced images, thus reducing the necessarycomputer resources. The metadata extraction is implemented with the detectionof a coronal dimming. To extract the coronal dimmings, we adopt the Reinardand Biesecker (2008) threshold, of more than 1σ below the mean pre-eventdifference image value. Figure 7 shows an example of our algorithm, in whichthe coronal dimmings associated with the 2000 Bastille Day event are extractedfrom the corresponding base difference data. Once the coronal dimming regionshave been identified, our algorithm analyses intensity changes over time usingnon-differenced data. Further metadata outputs of the algorithm include: area,lightcurves, location co-ordinates, volume and mass of the dimmings. Such out-puts can make an important contribution to the study of coronal dimmings andtheir ICME counterparts. This algorithm is discussed in more detail in Attrilland Wills-Davey (2009).

8. Jets

Extensive studies of polar coronal hole (CH) X-ray jets have been conductedwith the XRT aboard the Hinode satellite in early 2007 and late 2008. Coronalhole jets provide grounds for testing models of reconnection (Moreno-Insertis etal. 2009, in preparation) and may prove to be a source of the solar wind (Cirtainet al. 2007). Moreover, it was recently demonstrated by Savcheva et al. (2009,submitted to ApJL) that polar coronal hole jets are useful for understandinghow the solar cycle emerging flux polarities change. It has been suggested thatcontinuous monitoring of CH jets during the evolution of the solar cycle mayprove to be very useful for constraining certain solar dynamo models. Duringseveral days of observation in early 2007, 44 jets were identified in the 171ATRACE images. AIA, similar to TRACE in resolution and range of wavelengths,will have full-Sun field of view and consistent cadence. Therefore, AIA will beable to observe a great number of jets. The coronal jets found in coronal holesare the easiest to identify. Here, the coronal jet (CJ) stands for a jet inside acoronal hole.

The CJ detection calculations will take place at SAO, once triggered by botha bright point (BP) and CH identification by our system. The CH boundary



Figure 6. Left panel is a base difference image of EIT 195 A data (10:47 - 09:24 UT), showingthe coronal dimmings (black regions) during the 14 July 2000 Bastille Day event. White regionsindicate increased intensity compared to the pre-event (base) image at 09:24 UT. Regions thatdo not show any significant change in intensity with respect to the pre-event image appeargrey. Right panel shows the result of the coronal dimming algorithm. Only the dimming regionsare extracted.

is necessary to assure that the BP is inside the coronal hole. After identifying

the boundary of the coronal hole, the BP finder will be run every two images

and the pixels within BPs will be identified. The CJ detection and parameter

determination algorithms will work on data cubes covering a box enclosing

the BP and extending forward in time. Our methods for determining the CJ

parameters are described in detail in Savcheva et al. (2007). The detection

algorithm has already been tested on XRT images and testing on TRACE and

simulated AIA images will be conducted. In X-rays the BPs usually appear

on very smooth dark background of the CH Sun. In contrast, in EUV there is

much more detail that can be distinguished in the background and hence some

additional techniques will have to be employed to deal with this. The algorithm

is currently written in IDL. Processing each BP should not take more than 15s

on a 4-core 3-processor machine. The result from the algorithm is a VO Event

file containing the following VO Parameters1, a jet ID (KB ArchivID) and the

corresponding BP ID (KB ArchivID), the position of the origin of the jet in

heliospheric coordinates, a duration, length, width, line-of-sight velocity (km/s),

and BP size. The inclination with respect to the N-S direction and transverse

velocity are external parameters that will be saved in a text file with the other

parameters to create a metadata cube. A data cube with all images containing

the jet will also be produced. With the images metadata for each jet will occupy

about 90 MB.

1see http://www.lmsal.com/helio-informatics/hpkb/VOEvent Spec.html


Martens et al.

9. Oscillations

The study of large datasets (both in area and duration) is needed to improvethe statistics of current results regarding coronal oscillations as well as localised,short-lived events. Certain dynamical events such as CMEs, flares and coronaldimmings are known to be triggers for wave processes and hence, programsfor wave analysis will be run on data sets associated with such events, whichcomprise only a small fraction of the observations (< 0.1%). These codes arecurrently optimized for robustness and accuracy even at the expense of speed.Wavelet analysis will play an important role in this process, as it is capable ofidentifying not only periodicity present in a given data set, but can also provideboth spatial and time localisation (De Moortel and Hood 2000; De Moortel,Hood, and Ireland 2002; Ireland and De Moortel 2002). This ability to localise aperiodicity in time and space within a given dataset can be exploited to ‘train’the wavelet code to identify certain events.

The code we are adapting was originally developed by McAteer et al. (2004),and De Moortel and McAteer (2004). For a given datacube, it performs a waveletanalysis of the timeseries in each pixel and outputs the periodicities that arepresent, the duration (number of cycles) of the detected periodicities, as well astheir start and end time. These data will be entered into the VOEvent catalog.This information allows the team to optimize the wavelet code to detect avariety of events and features found in the solar atmosphere. Using such anautomated detection technique on specified observations (i.e. observations asso-ciated with CMEs, flares and coronal dimming events) will lead to a thoroughinvestigation into the link between dynamical (explosive) events and observedsolar oscillations. Indeed, De Moortel, Munday, and Hood (2004) showed thatvarious wavelet parameters can be altered to favor either the time or frequencyresolution. Additionally, by making a detailed study of the wavelet transform ofthe observed oscillations, it is possible to infer some coronal plasma properties,as demonstrated by De Moortel and Hood (2000), and De Moortel Hood, andIreland (2002).

The wavelet code is currently written entirely in IDL and is relatively CPU-intensive. The analysis of a datacube consisting of 150 200x300 pixel imagestakes roughly 60 minutes on an average desktop computer. We will explore avariety of options to improve performance for example by binning (which doesnot seem to affect wave detection). The code will be made available to users forsearches with user-specified optimization of search parameters.

10. EIT Wave Tracking

As part of the SDO computer vision project, EIT waves will be tracked and theirmetadata extracted using algorithms based on the work of Wills-Davey (2006)and Podladchikova and Berghmans (2005). Both sources rely on calibrated, dero-tated EUV data–Wills-Davey (2006) derive their output from TRACE 171 A and195 A percentage base difference images, while Podladchikova and Berghmans(2005) use data from the SoHO-EIT 195 A“CME Watch” and STEREO-EUVI



171 Aand 195 Adata. The Wills-Davey (2006) method assumes the EIT waveis a coherent front of intensity enhancement, and measures intensity increasesacross the field-of-view. Podladchikova and Berghmans (2005) instead find EITwaves by measuring the first four moments of a given full-disk EUV image; inparticular, the global kertosis offers a way to find EIT waves that may displacestructures while producing only dim fronts. This new algorithm takes aspects ofboth of these methods, enabling us to find a larger range of events.

Observations suggest that both EIT waves and coronal dimming regions arehighly correlated with CMEs, and many models assume these events are, in fact,a product of CME initiation, with EIT waves and dimming regions originatingcospatially. Because EIT waves are tracked using a polar coordinate system,the origin of these coordinates is determined by cross-referencing with the coro-nal dimming region tracking module, and finding the starting location of thedimming.

From here, two methods are used for finding EIT wave location. No assump-tions are made about the wave’s propagation characteristics, other than that itwill propagate roughly away from the flaring region. In cases where a bright frontis observed, the position of the intensity maximum is determined from one frameto the next in the polar coordinate frame. Figure 7 shows an example of thisautomated bright front tracking. This analysis is combined with measurements oflocal kertosis across overlapping 25 x 25-pixel sections. This use of local kertosisallows us to track subtle motion at a resolution of order 10”. Since EIT wavesare global structures, such a resolution is more than sufficient for determiningstructural displacement due to coherent wave propagation. Velocity is calculatedusing the same Huygens plotting technique discussed for the CME tracker.

In cases where intensity measurements show the existence of a bright front, werecord the intensity enhancements due to the wave as a function of time. Intensitycross-sections can either be calculated as a three-dimensional contoured array,or can be considered as “slices” along the Huygens-plotted velocity trajectories.As EIT waves are defined as propagating intensity increases, only the positivevalues from the difference images are kept. These data also also used to find thecenter-of-front amplitude as a function of time.

By combining intensity enhancement measurements from several overlappingEUV passbands, it is possible to measure the density increase above the pre-event background due to the wave front. Intensity measurements also allowus to calculate the entrained energy of a given wave event as follows: E =∫

d3V (P0∆ne2γne

−2γ) where E is the entrained energy, V is the volume ofinterest, P0 is the equilibrium pressure, ne is the electron density, and γ = 5/3.

All but the strongest EIT waves, while easy to pick out by eye in runningdifference movies, are nonetheless extremely difficult to find and track a priori,in that signal-to-noise tends to dominate the data; additionally, studies by Wills-Davey, Rachmeler, and Sechler (2008) show that false positive detections may bemore likely than previously expected. Because of this possibility of false positives,as well as the excessive resource requirements for wave detection from the fulldata stream, the wave detection program will be conducted post facto, triggeredby other events detected by our system, particularly flares and coronal dimmingregions. Output parameters will be input into the HPKB, using the appropriate


Martens et al.

Figure 7. Automated tracking of an EIT wave observed on 13 June 1998.

VOEvent format. We expect to observe 100-300 events/year, with disk space useof ∼ 2 Mb/event. Our codes are written in IDL and run on a single standarddesktop computer with short execution times.

11. Detection of Active Regions and Coronal Holes

The Spatial Possibilistic Clustering Algorithm (SPoCA) that we have developed(Barra et al., 2005; Barra, Delouille, and Hochedez, 2008; Barra et al., 2009)produces a segmentation of EUV solar images into regions named here ‘classes’corresponding to Active Regions (AR), Coronal Holes (CH), and the Quiet Sun(QS). SPoCA uses a multichannel fuzzy-logic clustering procedure. It has beenapplied successfully to a series of EIT image pairs (171 & 195 A) spanning almosta full solar cycle Barra et al. (2009). The classes are determined by minimizationof intra-class variance. The method is generic and therefore portable to otherinstruments, and in particular to AIA. SPoCA involves a preprocessing by whichthe limb brightness discontinuity is attenuated. Moreover, an initial calibrationstep determines the center values for each class, and thereby insures the temporalstability of the segmentation. SPoCA can take transformed EUV images as input,such as DEM maps obtained from AIA images (using software supplied by AIAinvestigators). This might help addressing the problem of line of sight confusion.

The level-1 product of the procedure is a set of maps giving the probability ofmembership to each class. Several higher level products can be generated from



these maps of probability:– A segmentation map, attributing a class to the pixel according to its highestprobability of membership.– A probability density function giving, for each pixel intensity value, the prob-ability that this pixel belong to a particular class.– Statistics such as area, mean intensity, and integrated intensity can be com-puted for each class.

From the maps of e.g. ARs, connected AR pixels are then gathered by way ofa region growing technique, see Figure 8. It provides the instantaneous locationof the center of mass of any given AR, its area, the coordinates of the boundingbox, and vectors of coordinates for all pixels belonging to the boundary (chaincode). These elements must be computed over time, and they will be handledas dynamical quantities. The current tracking method uses an optical flow al-gorithm to locate the ‘center of mass’ of the AR in the next fuzzy map, wherethe same region growing technique updates the parameters of the AR beingtracked. A starting date for an AR is defined when a new set of connected pixelsis identified. An AR ‘end-date’ is recorded either when the tracking algorithmcan no longer find a connected set, or when the AR disappears over the westlimb. The algorithm will also be able to deal with merging of ARs.

Figure 8. EIT image from 3 August 2000, together with overlays of segmented ARs. Anextracted AR is highlighted in white

Fuzzy CH maps can be treated exactly as the AR maps, producing area,location of ‘mass center’, and location of boundary for a connected element of theCH map. However, filament channels seen in coronal EUV passbands are often


Martens et al.

erroneously classified as ‘Coronal holes’ in the segmentation. Yet, exploiting allAIA channels, and particularly the ones sampling the Transition Region and thehot corona, should provide a solution to this equivocalness.

When Hα images are available, as is usually the case, the separation offilaments from CH becomes possible with the help of the filament detectionmodule described above, see Section 4.

On AIA images, we will identify AR, QS, and CH using our multichannelsegmentation algorithm on pairs of (171 & 195 A), (211 & 335 A) images, as wellas on 94 A images. Indeed, when similar morphological information is available intwo bandpasses, a multichannel approach provides a more robust segmentation,as compared to the monochannel case.

The procedure is written in C and currently takes 10 to 15 secondes to run ona Intel(R) Pentium(R) D CPU 2.80GHz when applied to a pair of 10242 images.

12. X-ray Bright Points

AIA will deliver a set of full frame images every ten seconds in each of its sevenEUV pass bands. The brightpoint detection algorithm will be run on the threecoolest coronal pass bands at wavelengths 171 A, 193 A, and 211 A.

We will use a modified version of the BP finder developed by McIntosh andGurman (2004) to sucessfully identify and track bright points in SoHO/EIT 171A, 193 AA, and 284 A over an entire solar cycle. The algorithm,written in IDL,uses full sun synoptic images to define areas of interest based on levels of intensityabove the background, and number of contiguous pixels that define shape andsize (see Fig. 9). This process is described in detail in Davey and McIntosh (2007;see also McIntosh and Gurman, 2005). The brightpoint detection algorithmworks by applying an NxN box car smoothing technique to determine the back-ground intensity. The noise level (σnoise) is also estimated. The background issubtracted from the original image and intensity enhancements (BP candidates)are classified by their separation from the background in units of σnoise, withsome restrictions on size and shape.

We plan several steps to improve the existing code:

- Include an upper and lower limit on the size of the BPs, weighted by limbangle.

Modify the region aspect ratio constraints to eliminate long ribbon-likestructures.

-- The McIntosh code can pick up strings of “BPs” which are actually thebrighter portions of longer ribbon/loop-like features (see Fig. 9). Afterinitial passes of BP detection, we will slightly reduce the noise thresholdto see if spatially nearby BPs merge. If so, the underlying ribbon structurewill then be reassigned.

- The McIntosh code can pick up groups of locally enhanced quiet Sun(QS) pixels, often with complex boundaries, which are clearly not BPs.We will place an additional constraint on the ratio of the two major axes



Figure 9. Features detected with automatic bright point finder algorithm in a SoHO/EIT171 A image on 01/03/2008 (left). The algorithm identifies bright points in EIT 171 A, 193A, and 284 A (middle), and with some simple modifications can also detect coronal holes andActive Regions (right).

of the BP and its perimeter/area ratio to prevent the detection of such QSconglomerations.

Once the bright points are defined by heliographic coordinates and time ofobservation, obtaining statistical information will be straightforward. The totalnumber, intensity (mean count rate), size, perimeter, major/minor axes, and areaof bright points will be determined for each of the regions in each bandpass. Todetermine lifetimes, the feature finding program will track the intensity-weightedcenter of mass and area in any of the AIA bassbands, with the position comparedwith projected rotation rates as a function of latitude. The lifetime of BPs willbe found from their appearance and disappearance, or their rotating on and offthe disk.

The output of the region finder stored in the HPKB will include time ofobservation, heliospheric coordinates, total area, integrated intensity for eachwavelength, and environment associated with each bright point (QS, AR, CH).Also associated with a select number of BPs will be a 128x128 image of thebright point region that can be made into quicktime movies, accessible, e.g.,through the VSO interface.


Martens et al.

Data processing time for finding and tracking bright points in three full diskEIT 171 A images is 4–5 seconds using a 2x3 GHz Dual-Core Intel Xeon Macin-tosh HD. Actual processing time will depend upon the AIA image preparationsoftware. This feature-finding algorithm will be ready for immediate applicationto the AIA data pipeline. Shortly after launch, detection parameters (back-ground, dark subtraction, sigma enhancement in each filter) will be adjusted afterin-flight calibrations and the AIA image preparation software becomes available.Algorithms that will correlate the feature finding database with other databaseswill be developed within the first year after launch depending upon when thelevel 2 algorithms are implemented. Metadata for BPs in each of the three fullframe images will be contained in text files that will occupy approximately 600kB per synoptic period, approximately every ten minutes. See Figure 13 for afull process flow chart of our BP algorithm.

Correlation with Other Databases. The heliographic coordinates deter-mined from the feature detection program will be correlated to other databaseswithin the pipeline to create other level 3 data products.The magnetic fluxtracker SWAMIS identifies and tracks magnetic flux concentrations above aparticular threshold (see Sec. 13). Using coordinates of bright points identifiedin the AIA 171 A channel, the magnetic flux database is cross-referenced forfootpoints within an area of the bright point coordinates. The magnetic flux foreach of the bright points can then be calculated by summing over the total areain each of the footpoints. The lifetime and unsigned magnetic flux of the regionwill be appended to the bright point catalog.

13. Magnetic Feature Tracking and Sunspots

We will provide comprehensive data on flux emergence, interaction, and can-cellation over the whole solar disk. Two main types of magnetic tracking dataare available: feature-level data, which represents the motion of resolved line-of-sight magnetic fields over the entire surface of the Sun within 45 degreesof the sub-Earth point; and large-scale emergence data, which identifies andhighlights emerging flux regions associated with ephemeral active regions on thesupergranular scale and above.

The feature-level data provide comprehensive motion and history informationon small magnetic details. They are intended to enable correlative statisticalstudies of flux origin, interaction with coronal features, and diffusion over thesolar surface. For example, it will be possible to extract and search on thesimplified magnetic geometry, and hence number of coronal null points andfield complexity, in the neighborhood of any feature or event in the entire SDOdata set. Further, every emergence event observed by SDO/HMI over the entiremission life will be cataloged and sorted by position, time, and event size andstrength, permitting statistical surveys of the interaction of emerging flux withexisting magnetic structures in the chromosphere and corona.

Magnetic feature tracking is easy to prototype but difficult to perform repro-ducibly and reliably. The pipeline uses a variant of the SWAMIS code describedby DeForest et al. (2007). For feature-level tracking, we use the ”downhill”



quasi-watershed method described by them; for large scale emergence trackingwe are prototyping both the multi-resolution method pioneered by Hagenaarand Cheung (2009) and a post-track feature clustering method; a final decisionwill be made during SDO commissioning. SWAMIS has been used to study fluxemergence patterns with SOHO/MDI (e.g. Lamb et al. 2008a) and Hinode/SOT-NFI (e.g. DeForest et al. 2008, Lamb et al. 2008b). Our basic tracking cadenceis six minutes; lifetime studies using SOHO/MDI data show that this is fastenough to capture evolution at the ∼1 arcsec resolution of SDO/HMI.

Simple feature location and evolution data are the most basic products ofevery feature recognition code. In addition, we provide origin and demise in-formation on each feature detected. SWAMIS classifies birth events into fivecategories: Isolated Appearances (a large fraction of new flux features appearin isolation, far from any other feature, as described by Lamb et al. 2008a);Fragmentations (in which a feature “calves” from a like-sign feature that isundergoing shredding); Emergences (in which a new feature appears in a man-ner that approximately conserves flux, either due to a new opposing featureappearing nearby or to growth of an existing nearby opposing feature); Errors(which do not appear to conserve flux); and Complex events where more thantwo features appear to be interacting directly. Under 1% of events greater than10× the detection threshold originate in Error events. In each case, SWAMISalso identifies the associated opposing region. Cancellation is classified using thesame scheme, reversed in time, to yield Disappearances, Mergers, Cancellations,Errors, and Complex cancellations.

Identification of every feature interaction event is crucial for the new field of”event-selected ensemble imaging” (ESEI), which allows deep-field study wellbelow the noise floor of an individual observation. ESEI using a combination ofmagnetic and other feature detections is a unique and powerful tool to distinguishmodels of small-scale activity that cannot be resolved any other way.

14. Polarity Inversion Line Mapping

Identifying the location of polarity inversion lines–for brevity, often called neutrallines–can be of great importance to phenomenological and theoretical studies.Historically, neutral lines in active regions have been useful tools for predictingthe locations of flares and coronal mass ejections Falconer, Moore, and Gary(2002). They also can be used to map out coronal structures McIntosh (1994)and are associated with filaments and filament channels (e.g. Martin et al. 1994;Chae et al. 2001) .

At present, little general work appears to have been done concerning the devel-opment of automated polarity inversion line mapping. Inversion lines have typi-cally either been drawn by hand or determined by contouring algorithms withinlarger programming packages. The use of a neutral line mapping algorithm notonly allows for user-independent inclusion in a pipeline, but standardizes themapping with pre-specified, explicit resolution scales.

To find polarity inversion lines, we will be using a well-established code previ-ously developed for NSO/KPVT magnetogram data (Jones 2004). This method


Martens et al.

invokes techniques such as edge detection, smoothing, and block averaging inorder to determine the appropriate scale at which to measure the neutral lines.This scale can be varied, allowing mapping to be done on a variety of lengthscales (Fig. 10). As neutral lines change slowly in the photosphere, the code willbe run on a ∼ 2 hour cadence on calibrated HMI vectormagnetogram data.

ss ss

Figure 10. Polarity inversion contours (shown in yellow) for length scales of 16” and 64”,respectively.

The Jones (2004) algorithm is written in IRAF. The output from this code willbe converted to VOEvent format as a mask data type. As there is only one outputparameter, and the associated methods are not CPU-intensive (requiring only astandard desktop computer), we estimate disk space requirements of 2 Mb/day.

15. Nonlinear Force-free Field Extrapolations

The solar magnetic field is key to understanding the physical processes in thesolar atmosphere. Unfortunately, we can measure the magnetic field vector rou-tinely with high accuracy only in the photosphere. These measurements areextrapolated into the corona under the assumption that the field is force-free,because the magnetic pressure is several orders of magnitudes higher than theplasma pressure. We have to solve the equations: (∇×B)×B = 0 and ∇·B = 0.

We solve by minimizing the functional proposed in Wheatland, Sturrock, andRoumeliotis (2000) and effectively encoded in cartesian and spherical geometryin Wiegelmann (2004)) and Wiegelmann (2007), respectively.

L =

∫

V

[

B−2 |(∇× B) × B|2 + |∇ ·B|2]

d3x, (1)

where an observed preprocessed vectormagnetogram specifies the photosphericboundary conditions. Preprocessing of the measured photospheric vector mag-netograms is necessary, because non-magnetic forces like plasma pressure andgravity are present in the photosphere and consequently the measured magnetic



field data are not consistent with force-free consistency criteria as defined in Aly(1989). We developed a minimization procedure that uses the measured pho-tospheric field vectors as input to approximate a more chromospheric-like fieldWiegelmann, Inhester, and Sakurai (2006). The procedure includes force-freeconsistency integrals and spatial smoothing. Direct chromospheric observationscan be taken into account for an improved match to the field direction as inferredfrom Hα fibrils by Wiegelmann et al. (2008).

A recent comparison of different nonlinear force-free extrapolation codes re-vealed that minimizing the functional (1) is the fastest and most accurate cur-rently available method (Schrijver et al., 2006; Metcalf et al., 2008). Computingthe force-free magnetic field in a 3003 cube takes about 2h on a four processormachine.

We intend to use the large and high resolution FOV-vectormagnetogramsprovided by SDO/HMI to produce several times per day estimates for the freemagnetic energy for the ARs on disc, and produce 3D magnetic field mapsoutlining the general coronal magnetic field topology. These data will be cross-referenced with the AR and CH data obtained through other methods. (SeeSect. 11).

Further complications in using current vector magnetograms (e.g. from Hin-ode/SOT) are a limited field of view, missing data and high noise in the trans-verse B-field measurements (DeRosa et al., 2009). For meaningful extrapolationsit is important to use large and high resolution FOV-vectormagnetograms - asprovided by SDO/HMI - and to deal with measurement errors and non-magneticforces by pre-processing. For reliable estimation of the free magnetic energy andtopology from the extrapolated coronal magnetic field model, one should validatethe model field by additional coronal observations, e.g., compare projected fieldlines with SDO/AIA images.

16. Trainable Feature Recognition and Retrieval

Humans have an amazing generic feature recognition ability that has been hardto match for computers. For example, a solar scientist can instruct an undergrad-uate student in less than an hour to recognize sunspots, filaments, loops, andarcades in solar imagery, and the student can then easily produce a catalog ofthese features from a given set of images. A computer feature finding algorithmon the other hand takes months to years to develop, and the development needsto be repeated almost from scratch for every new feature.

Motivated by the successful development and implementation of a generic,more human-like, feature detection method for mammography (e.g. Yang etal., 2007) we are in the process of creating an automated solar feature retrievalsystem that is generic in nature, i.e., the software can detect features of any kind,and rank the returned images based on their similarity to an image provided bythe user (see Fig. 11 for a sample of the user interface). Rather than developing anew task-specific application to identify each separate feature, our generic featurerecognition software can detect and catalog a wide range of solar features, eveninvolving serendipitous discovery. This work will benefit the solar community


Martens et al.

in three ways by delivering: (1) a benchmark data set, that includes balancedrepresentations of common types of phenomena, and can be use to performcomparative evaluation of different image recognition systems, which will bemade freely available for everyone to use, (2) catalogs for common types ofphenomena with database indexes, speeding up the data search and retrieval, (3)a content-based image retrieval system with a ranking mechanism, that returnsimages based on their similarity to the image provided as a query, and can learnfrom user’s feedback.

Figure 11. An example of the querying interface showing sample results with the option torate the returned images and resubmit the image-based query.

Method. There are two steps to our method. The first is implemented atthe end of the metadata pipeline, and the second is totally separate from it andcan be performed by any user at any time, even on her or his laptop, withoutrequiring resources from the SDO JSOC. Step one consists of calculating foreach AIA image the texture parameters that are then stored in the image tex-ture (meta-)catalog. Specifically, each AIA image is subdivided in 1024 128X128sections and for each of those a set of texture parameters such as entropy, averageintensity, standard deviation, kurtosis, etc. is calculated. The parameters that weare currently using for analyzing TRACE images are given in Table 2. Assumingthat we derive 16 parameters per image section, each 16 bits, the informationcompression from AIA image to meta-catalog entry is roughly a factor 1000.Therefore the size of the AIA image meta-catalog will not be prohibitive. Weestimate that the calculation of statistics for a single image entry into the catalogrequires one sec. of computing time for the whole grid-segmented image.



Table 2. Texture parameters currently used for analyz-ing TRACE images.

Name & Equation

Mean m = 1L

∑

L−1

i=0zi

Standard Deviation σ =

√

1L

∑

L−1

i=0(zi − m)2

Third Moment µ3 =∑

L−1

i=0(zi − m)3p(zi)

Fourth Moment µ4 =∑

L−1

i=0(zi − m)4p(zi)

Relative Smoothness R = 1 − 11+σ2(z)

Entropy E = −∑

L−1

i−0p(zi) log2 p(zi)

Uniformity U =∑

L−1

i=0p2(zi)

Suppose now that a user wants to build a catalog of loop arcades by teachingthe algorithm how to detect them. That user then downloads a selection, sayseveral dozens, of AIA full disk images. In those images the user identifies arcadesvia the simple point-and-click interface shown in Figure 11; there is no needto identify all arcades on the disk. The program then calculates the textureparameters for each of the image segments that contain the identified arcades.Using these parameters as a feature definition, the entire AIA image texturemeta-catalog can be quickly searched to filter out irrelevant images and thenfine tune the search for image segments with similar parameters. These segmentswill contain arcades if the image texture parameters are adequate for the typeof image one is analyzing.

Our work on TRACE images (Lamb 2008; Lamb, Angryk, and Martens 2008)so far has focused on the crucial step of determining the correct texture parame-ters (those that have good properties, first from the perspective of computationalcosts to keep up with our pipeline, and second from the perspective of distin-guishing between different types of phenomena). Table 3 summarizes our currentresults. Here the Receiver Operating Characteristic (ROC) curve plots the true-positive rate on the y-axis and the false-positive rate on the x-axis. The areaunder the curve represents the accuracy of the classifier overall. The closer thearea under the curve is to 1.0, the more accurate the classifier, while at 0.5 theclassifier would not be any better than randomly picking a class. Random UnderSampling (RUS) and Random Over Sampling (ROS) are different samplingmethods to treat data sets with unbalanced numbers of different phenomena(e.g., relatively common images of quiet sun and coronal loops, versus relativelyrare occurrences of flares and filaments). The column headers represent differ-ent classification techniques (C4.5 performs entropy-based classification, SVMstands for Support Vector Machines, and AdaBoost is a commonly used boostingtechnique, that emphasizes misclassified samples during classifier training).


Martens et al.

At the current moment, we continue to evaluate our texture parameters andimprove our classifiers using TRACE images. We anticipate that the TRACEresults will carry over easily to AIA because of the similarity of their imagery.We emphasize here again that step two is completely separate from the datapipeline, and that the AIA data archive need not be accessed in its entirety, onlya small number of images needs to be downloaded to provide a training sampleto our software. Since the algorithm is further trainable, a user can fine-tune theselection, e.g. to distinguish between left-skewed and right-skewed arcades.

Table 3. Average Area Under an ROC curve using two different sampling techniques(RUS and ROS), and two different classification algorithms (C4.5 and SVM) fordifferent types of solar phenomena.

C4.5 SVM AdaBoost C4.5 AdaBoost SVM

Phenomenon RUS ROS RUS ROS RUS ROS RUS ROS

Quiet Sun 0.795 0.872 0.940 0.923 0.920 0.912 0.939 0.920

Coronal Loop 0.897 0.905 0.932 0.922 0.911 0.917 0.912 0.912

Sunspot 0.890 0.917 0.922 0.958 0.901 0.932 0.944 0.943

Filament 0.838 0.960 0.832 0.848 0.875 0.783 0.898 0.897

Flare 0.977 0.970 0.988 0.980 0.977 0.967 0.977 0.976

Our current list of features that we intend to generate VOEvent entries for(time and location at a minimum), using this method, includes: cusps, arcades,null-geometries, flare ribbons (starting with the flare catalog), keyholes (previ-ously detected by Skylab and Yohkoh-SXT), circular filaments (starting withthe filament catalog), faculae, pores, surges, arch filaments, delta-spots (frommagnetograms), plumes,and anemones (also called rosettes).

Discovery of New Features. A promising advantage of generic featurerecognition methods over task-specific ones is the potential for rapid discoveryand cataloging, even serendipitously, of new features. An example is given inFig. 12, a TRACE image of an Active Region having a special type of loopgeometry. The user can start with this image to try to find similar ones inthe TRACE or AIA meta-catalog. Not many such images have been seen withTRACE, probably because they are indeed rare, and for an instrument likeAIA, searching the database to find more of such images is a daunting task. Ourimage texture characterization makes such a search possible and even rathersimple without the user needing to know how to specify the parameters of thesearch. The scientific interest of Fig. 12 is clear: this is what one would expectto see in the neighborhood of a magnetic null, making it a good candidate forinclusion in a study of such topologies.

This then points to a more general application for the image texture meta-catalog. A user may see a number if images that look very similar and which



Figure 12. TRACE 171 A AR image showing a loop geometry that is consistent with thepresence of a magnetic null.

pique interest without even being able to define exactly what is exceptional

about these images (Fig. 12 would be an example for someone with limited

knowledge of MHD). Our feature recognition method can then be used to findsimilar images without an excessive effort, even though the user does not even

know exactly what he or she is looking for!

The image texture (meta)-catalog, together with a manual and supporting

software, will be made available on-line separately from the VOEvent catalog,because the structures are incompatible.

17. A Pipeline for the Generation of Level 3 Metadata

Computing Facilities at LMSAL. LMSAL will provide infrastructure to

enable the running of appropriate feature and event modules. This includes a

cluster of Linux servers, running SUSE Linux, and consisting of either 4 or 8core 2.2GHz processors with up to 16Gb of memory per server. LMSAL has

available >100Tb of storage that will be NFS-mounted to these servers. To

enable fast communication between nodes, infiniband connections will be used.

Data transfer between Stanford and LMSAL is achieved by means of a 10Gbpsfiber link, giving LMSAL almost-instant access to the SDO data. As part of

their commitment to implementing the HPKB (described in a companion paper

in this issue), LMSAL provides codes in IDL for interacting with DRMS/SUMSand extracting data needed by the feature and event detection modules. LMSAL

will also provide a control system for running and monitoring these modules and

tools that enable the easy contribution of events to the HPKB. In the AIA Level


Martens et al.

2 Design Document2, LMSAL describe the Event Identification Sytem (EIS),which continually acquires incoming data and makes it available to a variety offeature and event modules.

Because the available resources on which to run the various feature and eventdetection modules are finite, SAO will provide all the additional computingpower required to successfuly run the modules

SAO Computing Facilities. SAO has recently deployed an 11-node com-pute cluster consisting of 2 dual core 3 GHz processors with between 16 and32 Gb of memory per node. It has also procured a 280 TB storage array fromSUN. The storage array will be available to nodes via NFS, which has shownto be more than capable of delivering data at the required rates. SAO will usethe storage array to construct a rolling archive of the most recent months ofSDO HMI and AIA 1.5-level data, excluding the helioseismology data. SAO hasalready implemented and tested a remote version of the DRMS/SUMS setupthat is in use at Stanford. Stanford will be using SLONY-I5 (a master-slavereplication system that replicates large databases to a number of slave systems),to replicate meta-information to a limited number of external sites, includingSAO. SAO will then transfer and store SDO data in the local SUMS system asit becomes available. This will give SAO rapid access to SDO data and enablelocal processing of SDO data, and easy comparison between SDO data and datafrom other sources that are processed locally (such as H-α data in the lamentdetection module). Interfacing to DRMS. LMSAL will provide a dedicated IDLbased interface to DRMS for use by IDL-based feature and event detectionmodules. SAO will also make use of code locally for any IDL module processedthere. For non-IDL modules, e.g. the magnetic feature tracking module, writtenin PDL, the code will interface directly with DRMS.

SAO will run algorithms that either detect events not relevant to SDO/HPKBeffort at LMSAL or report exhaustive findings (e.g. x-ray bright point locations),and so the SAO computer vision center will port the EDS software to one of theirclusters. SAO will also run detection algorithms that do not use SDO data thatinclude filaments and CMEs as well as modules that key off of detection eventscreated by other detection algorithms (e.g. flare detection events will spawnthe process to look for oscillations). At least initially, there will be a need forhuman interaction for some event types, EIT waves and oscillations, to ensurethe detections represent actual events. A detailed rendition of the data flow inthe case of the X-ray bight point detection module is given in Figure 13.

Module Testing. The facilities available at SAO will enable us to create aframework similar to that available for running the feature and event detectionmodules at LMSAL. We will utilize this framework for end-to-end testing andevaluation of the modules before they are deployed at LMSAL, and for prototyp-ing algorithmic or coding improvements to the modules. Part of the evaluationwill include investigation of standardization opportunities, particularly in thecase of the IDL codes, which will have to be modified, to interact with DRMSand provide output in VOEvent format, using the environment provided by

2http://vso.stanford.edu/hekwiki/FrontPage?action=AttachFile&do=get&target=AIAEventOverview.pdf



JSOCHead Node

RAMFS

L1 files sent

to BP

algorithm

node

XBP

identified in

single image

Text file

retained in

SATA

disk on node #n

Text file of Daily record of BP

locations for 93Achannel updated

copy of L1 file sent to node 'n' for XBP processing

Process repeated for every tenth image

collected during UT day

Text file for Daily XBP

locations/durations/ etc

converted into

VOEvent file

Metadata

Catalogue on

Archive SATA

Daily XBP file for

each AIA filter

used to

determine

space/time

correlations for

XBPs

NOTE: one VOEvent file is created for each unique

XBP

QuickLook

Product for XBP

After VOEvent creation, a thumbnail image for the

event is created

AIA identified XBPs are

compared to XRT

identified XBPs,

VOEvent file updated

*Final* VOEvent file used to populated metadata catalogue

Pointer to QL

product created

for VSO database

QL data available via

VSO; metadata

catalogue search

tools available

through VSO portal

NOTE: QL data products are retained on-disk for n-months (TBD). "On the fly"

QL products can be created for n+-month events

For file requests older than a-months, an SAO provided interface to the JSOC

archive (DRMS-compliant) will retrieve the files and deliver them to the client. This

may require an extended wait-time (TBR).

Metadata Capture Process for X-ray Bright Point

Figure 13. Pipeline flow-chart for image data (input) and meta-data (output) in the case ofthe X-ray bright point module.

LMSAL. This will include utilizing the procedures being developed at LMSAL

for creating and accessing shared memory segments between multiple processes,

and file-memory mapping for the faster input of image files.

Community Access to Metadata at SAO. The primary target for the

feature and event detection modules’ output is the SDO/JSOC feature metadata

catalog (i.e., HPKB at LMSAL). In addition to primarily supplying the HPKB

with features and events, SAO will utilize the newly redesigned VSO catalog

infrastructure to make features and events available, with search capabilities

that extend beyond those currently proposed for the HPKB. This will include

the ability to use the catalogs as datasets for doing science. In addition, the VSO

catalog infrastructure will allow users to comment on data and catalog entries.

This information can be used to enhance the data being provided. Additionally,

some of the feature and event detection modules will be run at SAO (e.g. for

non-SDO data) requiring SAO to host any secondary data products generated

as the module runs. By utilizing the VSO infrastructure, LMSAL and SAO can

easily make data available to other VOs and provide enhanced integration for all

of Heliophysics. Dissemination of the pipeline capabilities will be accomplished

through the use of a guidebook, two annual workshops, and a web interface.


Martens et al.

18. Conclusions

This infrastructure development project enables otherwise inaccessible science byallowing unprecedented levels of feature selection and cross-correlation. Fifteenseparate feature and event detection modules, plus a trainable feature recognitionpostprocessing subsystem, provide comprehensive indexing of every currentlyaccessible solar feature using current computer vision techniques; our modu-lar infrastructure permits simple addition of future modules as needed and/oravailable.

It is difficult to overstate the importance of comprehensive cross-feature in-dexing with uniform, modular systems that are vetted for operational use. Forexample, the new field of event-selected ensemble imaging (Lamb et al., 2008) hasbeen used to characterize the small-scale dynamo with magnetic feature track-ing; applied to cross-feature selections, Event-Selected Ensemble Imaging (ESEI)and similar techniques will enable systematic studies of myriad phenomena thathitherto have been studied anecdotally.

Acknowledgements We thank Dr. Trae Winter for carefully proofreading the manuscript.

References

Aly, J.J.: 1989, On the reconstruction of the nonlinear force-free coronalmagnetic field from boundary data. Sol. Phys. 120, 19 – 48.

Asai, A., Ishii, T.T., Kurokawa, H., Yokoyama, T., Shimojo, M.: 2003, Evolutionof Conjugate Footpoints inside Flare Ribbons during a Great Two-RibbonFlare on 2001 April 10. Astrophys. J. 586, 624 – 629. doi:10.1086/367694.

Attrill, G.D.R., Wills-Davey, M.: 2009, Automatic detection and extrac-tion of coronal dimmings for SDO/AIA data for implementation in theSDO/Computer Vision Center. Solar Phys. submitted.

Barra, V., Delouille, V., Hochedez, J.F.: 2008, Segmentation of extreme ultravio-let solar images via multichannel fuzzy clustering. Advances in Space Research42, 917 – 925. doi:10.1016/j.asr.2007.10.021.

Barra, V., Delouille, V., Hochedez, J.F., Chainais, P.: 2005, Segmentation of EITImages Using Fuzzy Clustering: a Preliminary Study. In: The Dynamic Sun:Challenges for Theory and Observations, ESA Special Publication 600.

Barra, V., Delouille, V., M., K., Hochedez, J.F.: 2009, Fast and robust segmen-tation of solar EUV images: algorithm and results for solar cycle 23. A in

press.

Bemporad, A., Raymond, J., Poletto, G., Romoli, M.: 2007, A ComprehensiveStudy of the Initiation and Early Evolution of a Coronal Mass Ejectionfrom Ultraviolet and White-Light Data. Astrophys. J. 655, 576 – 590.doi:10.1086/509569.



Bernasconi, P.N., Georgoulis, M.K.: 2009, . Solar Phys. this volume.

Bernasconi, P.N., Rust, D.M., Hakim, D.: 2005, Advanced Automated SolarFilament Detection And Characterization Code: Description, Performance,And Results. Solar Phys. 228, 97 – 117. doi:10.1007/s11207-005-2766-y.

Bewsher, D., Harrison, R.A., Brown, D.S.: 2008, The relationship between EUVdimming and coronal mass ejections. I. Statistical study and probabilitymodel. Astron. Astrophys. 478, 897 – 906. doi:10.1051/0004-6361:20078615.

Canfield, R.C., Hudson, H.S., McKenzie, D.E.: 1999, Sigmoidal morphol-ogy and eruptive solar activity. Geophys. Res. Lett. 26, 627 – 630.doi:10.1029/1999GL900105.

Canfield, R.C., Kazachenko, M.D., Acton, L.W., Mackay, D.H., Son, J., Free-man, T.L.: 2007, Yohkoh SXT Full-Resolution Observations of Sigmoids:Structure, Formation, and Eruption. Astrophys. J. Lett. 671, 81 – 84.doi:10.1086/524729.

Chae, J., Martin, S.F., Yun, H.S., Kim, J., Lee, S., Goode, P.R., Spirock, T.,Wang, H.: 2001, Small Magnetic Bipoles Emerging in a Filament Channel.Astrophys. J. 548, 497 – 507. doi:10.1086/318661.

Christe, S., Hannah, I.G., Krucker, S., McTiernan, J., Lin, R.P.: 2008, RHESSIMicroflare Statistics. I. Flare-Finding and Frequency Distributions. Astro-phys. J. 677, 1385 – 1394. doi:10.1086/529011.

Cirtain, J.W., Golub, L., Lundquist, L., van Ballegooijen, A., Savcheva, A.,Shimojo, M., DeLuca, E., Tsuneta, S., Sakao, T., Reeves, K., Weber, M.,Kano, R., Narukage, N., Shibasaki, K.: 2007, Evidence for Alfven Waves inSolar X-ray Jets. Science 318, 1580. doi:10.1126/science.1147050.

Davey, A.R., McIntosh, S.: 2007, The SoHO/EIT Brightpoint Database: Min-ing The Database For Science. In: American Astronomical Society MeetingAbstracts, American Astronomical Society Meeting Abstracts 210, 95 – 09.

de Moortel, I.: 2005, An overview of coronal seismology. Royal Society of LondonPhilosophical Transactions Series A 363, 2743 – 2760.

De Moortel, I., Hood, A.W.: 2000, Wavelet analysis and the determination ofcoronal plasma properties. Astron. Astrophys. 363, 269 – 278.

De Moortel, I., McAteer, R.T.J.: 2004, Waves and wavelets: An auto-mated detection technique for solar oscillations. Solar Phys. 223, 1 – 2.doi:10.1007/s11207-004-0806-7.

De Moortel, I., Hood, A.W., Ireland, J.: 2002, Coronal seismology throughwavelet analysis. Astron. Astrophys. 381, 311 – 323. doi:10.1051/0004-6361:20011659.


Martens et al.

De Moortel, I., Munday, S.A., Hood, A.W.: 2004, Wavelet Analysis: theeffect of varying basic wavelet parameters. Solar Phys. 222, 203 – 228.doi:10.1023/B:SOLA.0000043578.01201.2d.

DeForest, C.E., Plunkett, S.P., Andrews, M.D.: 2001, Observation of Po-lar Plumes at High Solar Altitudes. Astrophys. J. 546, 569 – 575.doi:10.1086/318221.

DeForest, C.E., Hagenaar, H.J., Lamb, D.A., Parnell, C.E., Welsch, B.T.:2007, Solar Magnetic Tracking. I. Software Comparison and RecommendedPractices. Astrophys. J. 666, 576 – 587. doi:10.1086/518994.

DeForest, C.E., Berger, T.E., Katsukawa, Y., Lamb, D.A., J., H.H., Parnell,C.P., Welsch, B.T.: 2008, Where is the small-scale mixed-polarity flux? In:AGU Joint Assembly Abstracts.

DeRosa, M.L., Schrijver, C.J., Barnes, G., Leka, K.D., Lites, B.W., Aschwan-den, M.J., Amari, T., Canou, A., McTiernan, J.M., Regnier, S., Thalmann,J.K., Valori, G., Wheatland, M.S., Wiegelmann, T., Cheung, M.C.M., Conlon,P.A., Fuhrmann, M., Inhester, B., Tadesse, T.: 2009, A Critical Assessmentof Nonlinear Force-Free Field Modeling of the Solar Corona for Active Region10953. ApJ 696, 1780 – 1791. doi:10.1088/0004-637X/696/2/1780.

Falconer, D.A., Moore, R.L., Gary, G.A.: 2002, Correlation of the Coronal MassEjection Productivity of Solar Active Regions with Measures of Their GlobalNonpotentiality from Vector Magnetograms: Baseline Results. Astrophys. J.569, 1016 – 1025. doi:10.1086/339161.

Fan, Y., Gibson, S.E.: 2003, The Emergence of a Twisted Magnetic FluxTube into a Preexisting Coronal Arcade. Astrophys. J. Lett. 589, 105 – 108.doi:10.1086/375834.

Fan, Y., Gibson, S.E.: 2004, Numerical Simulations of Three-dimensional Coro-nal Magnetic Fields Resulting from the Emergence of Twisted Magnetic FluxTubes. Astrophys. J. 609, 1123 – 1133. doi:10.1086/421238.

Fisher, R.R., Munro, R.H.: 1984, Coronal transient geometry. I - Theflare-associated event of 1981 March 25. Astrophys. J. 280, 428 – 439.doi:10.1086/162009.

Gibson, S.E., Fan, Y., Torok, T., Kliem, B.: 2006, The Evolving Sigmoid: Ev-idence for Magnetic Flux Ropes in the Corona Before, During, and AfterCMES. Space Science Reviews 124, 131 – 144. doi:10.1007/s11214-006-9101-2.

Gilbert, H.R., Holzer, T.E., Burkepile, J.T., Hundhausen, A.J.: 2000, Activeand Eruptive Prominences and Their Relationship to Coronal Mass Ejections.Astrophys. J. 537, 503 – 515. doi:10.1086/309030.



Golub, L., Deluca, E., Austin, G., Bookbinder, J., Caldwell, D., Cheimets, P.,Cirtain, J., Cosmo, M., Reid, P., Sette, A., Weber, M., Sakao, T., Kano, R.,Shibasaki, K., Hara, H., Tsuneta, S., Kumagai, K., Tamura, T., Shimojo, M.,McCracken, J., Carpenter, J., Haight, H., Siler, R., Wright, E., Tucker, J.,Rutledge, H., Barbera, M., Peres, G., Varisco, S.: 2007, The X-Ray Telescope(XRT) for the Hinode Mission. Solar Phys. 243, 63 – 86. doi:10.1007/s11207-007-0182-1.

Green, L.M., Kliem, B., Torok, T., van Driel-Gesztelyi, L., Attrill, G.D.R.: 2007,Transient Coronal Sigmoids and Rotating Erupting Flux Ropes. Solar Phys.246, 365 – 391. doi:10.1007/s11207-007-9061-z.

Hagenaar, H.: 2009, Magnetic Flux Emergence on Different Scales. Hinode 2proceedings, submitted.

Ireland, J., De Moortel, I.: 2002, Application of wavelet analysis to transversalcoronal loop oscillations. Astron. Astrophys. 391, 339 – 351. doi:10.1051/0004-6361:20020643.

Jing, J., Yurchyshyn, V.B., Yang, G., Xu, Y., Wang, H.: 2004, On the Relationbetween Filament Eruptions, Flares, and Coronal Mass Ejections. Astrophys.J. 614, 1054 – 1062. doi:10.1086/423781.

Jones, H.P.: 2004, ”counting magnetic bipoles on the sun by polarity inversion”.In: ”Knowledge-Based Intelligent Information and Engineering Systems: 8thInternational Conference”, 433 – 439.

Kliem, B., Titov, V.S., Torok, T.: 2004, Formation of current sheets and sig-moidal structure by the kink instability of a magnetic loop. Astron. Astrophys.413, 23 – 26. doi:10.1051/0004-6361:20031690.

LaBonte, B.J., Rust, D.M., Bernasconi, P.N.: 2003, An Automated Systemfor Detecting Sigmoids in Solar X-ray Images. In: Bulletin of the AmericanAstronomical Society, Bulletin of the American Astronomical Society 35, 814.

Lamb, D.A., DeForest, C.E., Hagenaar, H.J., Parnell, C.P., Welsch, B.T.: 2008,Does network magnetic flux influence the small scale dynamo? In: AGU JointAssembly Abstracts.

Lamb, D.A., DeForest, C.E., Hagenaar, H.J., Parnell, C.E., Welsch, B.T.: 2008,Solar Magnetic Tracking. II. The Apparent Unipolar Origin of Quiet-Sun Flux.Astrophys. J. 674, 520 – 529. doi:10.1086/524372.

Lamb, R., Angryk, R., Martens, P.: 2008, An Example-based Image RetrievalSystem for the TRACE Repository. In: Proceedings of the 19th InternationalConference on Pattern Recognition ICPR 08.

Lamb, R.: 2008, An Information Retrieval System for Images from the TRACESatellite (URL=http://www.cs.montana.edu/techreports/2008/Lamb.pdf).Master’s thesis, Montana State University, Bozeman, MT, USA.


Martens et al.

Leblanc, Y., Dulk, G.A., Vourlidas, A., Bougeret, J.L.: 2001, Tracing shock wavesfrom the corona to 1 AU: Type II radio emission and relationship with CMEs.J. Geophys. Res. 106, 25301– 25312. doi:10.1029/2000JA000260.

Martin, S.F.: 1998, Conditions for the Formation and Maintenance of Filaments(Invited Review). Solar Phys. 182, 107 – 137. doi:10.1023/A:1005026814076.

Martin, S.F., Bilimoria, R., Tracadas, P.W.: 1994, Magnetic field configurationsbasic to filament channels and filaments. In: Rutten, R.J., Schrijver, C.J. (eds.)Solar Surface Magnetism, 303.

Masuda, S., Kosugi, T., Hara, H., Tsuneta, S., Ogawara, Y.: 1994, A Loop-Top Hard X-Ray Source in a Compact Solar Flare as Evidence for MagneticReconnection. Nature 371, 495. doi:10.1038/371495a0.

McAteer, R.T.J., Gallagher, P.T., Bloomfield, D.S., Williams, D.R., Math-ioudakis, M., Keenan, F.P.: 2004, Ultraviolet Oscillations in the Chromosphereof the Quiet Sun. Astrophys. J. 602, 436 – 445. doi:10.1086/380835.

McIntosh, P.S.: 1994, YOHKOH X-Ray Image Interpretation with Overlays ofHα Neutral Lines. In: Uchida, Y., Watanabe, T., Shibata, K., Hudson, H.S.(eds.) X-ray solar physics from Yohkoh.

McIntosh, S.W., Gurman, J.B.: 2004, EIT EUV Brightpoints Over the SOHOMission so Far. In: Walsh, R.W., Ireland, J., Danesy, D., Fleck, B. (eds.)SOHO 15 Coronal Heating, ESA Special Publication 575, 235.

McIntosh, S.W., Gurman, J.B.: 2005, Nine Years Of Euv Bright Points. SolarPhys. 228, 285 – 299. doi:10.1007/s11207-005-4725-z.

Metcalf, T.R., Derosa, M.L., Schrijver, C.J., Barnes, G., van Ballegooijen,A.A., Wiegelmann, T., Wheatland, M.S., Valori, G., McTtiernan, J.M.: 2008,Nonlinear Force-Free Modeling of Coronal Magnetic Fields. II. Modeling a Fil-ament Arcade and Simulated Chromospheric and Photospheric Vector Fields.Solar Phys. 247, 269 – 299. doi:10.1007/s11207-007-9110-7.

Micha lek, G., Gopalswamy, N., Yashiro, S.: 2003, A New Method for EstimatingWidths, Velocities, and Source Location of Halo Coronal Mass Ejections.Astrophys. J. 584, 472 – 478. doi:10.1086/345526.

Moreno-Insertis, F., Galsgaard, K., Ugarte-Urra, I.: 2009, X-ray Jets in CoronalHoles: Numerical Simulation and Hinode Observations. Hinode 2 proceedings,submitted.

Olmedo, O., Zhang, J., Wechsler, H., Poland, A., Borne, K.: 2008, Automatic De-tection and Tracking of Coronal Mass Ejections in Coronagraph Time Series.Solar Phys. 248, 485 – 499. doi:10.1007/s11207-007-9104-5.

Pevtsov, A.A., Balasubramaniam, K.S., Rogers, J.W.: 2003, Chirality of Chro-mospheric Filaments. Astrophys. J. 595, 500 – 505. doi:10.1086/377339.



Podladchikova, O., Berghmans, D.: 2005, Automated Detection Of Eit WavesAnd Dimmings. Solar Phys. 228, 265 – 284. doi:10.1007/s11207-005-5373-z.

Reinard, A.A., Biesecker, D.A.: 2008, Coronal Mass Ejection-Associated CoronalDimmings. Astrophys. J. 674, 576 – 585. doi:10.1086/525269.

Robbrecht, E., Berghmans, D.: 2004, Automated recognition of coronal massejections (CMEs) in near-real-time data. Astron. Astrophys. 425, 1097 – 1106.doi:10.1051/0004-6361:20041302.

Rust, D.M., Hildner, E.: 1976, Expansion of an X-ray coronal arch into the outercorona. Solar Phys. 48, 381 – 387.

Rust, D.M., Kumar, A.: 1996, Evidence for Helically Kinked Magnetic FluxRopes in Solar Eruptions. Astrophys. J. Lett. 464, 199. doi:10.1086/310118.

Rust, D.M., LaBonte, B.J.: 2005, Observational Evidence of the Kink Instabilityin Solar Filament Eruptions and Sigmoids. Astrophys. J. Lett. 622, 69 – 72.doi:10.1086/429379.

Rust, D.M., Martin, S.F.: 1994, A Correlation Between Sunspot Whirls andFilament Type. In: Balasubramaniam, K.S., Simon, G.W. (eds.) Solar ActiveRegion Evolution: Comparing Models with Observations, Astronomical Societyof the Pacific Conference Series 68, 337.

Rust, D.M., Anderson, B.J., Andrews, M.D., Acuna, M.H., Russell, C.T., Schuck,P.W., Mulligan, T.: 2005, Comparison of Interplanetary Disturbances at theNEAR Spacecraft with Coronal Mass Ejections at the Sun. Astrophys. J.621, 524 – 536. doi:10.1086/427401.

Savcheva, A., Cirtain, J., Deluca, E.E., Lundquist, L.L., Golub, L., Weber, M.,Shimojo, M., Shibasaki, K., Sakao, T., Narukage, N., Tsuneta, S., Kano, R.:2007, A Study of Polar Jet Parameters Based on Hinode XRT Observations.Pub. Astron. Soc. Japan 59, 771.

Savcheva, A., Cirtain, J., DeLuca, E., Golub, L.: 2009, Does a Polar CoronalHole Flux Emergence Follow a Hale-like Law? ApJL, submitted.

Schrijver, C.J., Derosa, M.L., Metcalf, T.R., Liu, Y., McTiernan, J., Regnier,S., Valori, G., Wheatland, M.S., Wiegelmann, T.: 2006, Nonlinear Force-FreeModeling of Coronal Magnetic Fields Part I: A Quantitative Comparison ofMethods. Solar Phys. 235, 161 – 190. doi:10.1007/s11207-006-0068-7.

Sheeley, N.R., Walters, J.H., Wang, Y.M., Howard, R.A.: 1999, Continuoustracking of coronal outflows: Two kinds of coronal mass ejections. J. Geophys.Res. 104, 24739 – 24768. doi:10.1029/1999JA900308.

Titov, V.S., Demoulin, P.: 1999, Basic topology of twisted magnetic configura-tions in solar flares. Astron. Astrophys. 351, 707 – 720.


Martens et al.

Torok, T., Kliem, B.: 2005, Confined and Ejective Eruptions of Kink-unstableFlux Ropes. Astrophys. J. Lett. 630, 97 – 100. doi:10.1086/462412.

Torok, T., Kliem, B.: 2007, Numerical simulations of fast and slow coronal massejections. Astronomische Nachrichten 328, 743. doi:10.1002/asna.200710795.

Tsuneta, S., Hara, H., Shimizu, T., Acton, L.W., Strong, K.T., Hudson, H.S.,Ogawara, Y.: 1992, Observation of a solar flare at the limb with the YOHKOHSoft X-ray Telescope. Pub. Astron. Soc. Japan 44, 63 – 69.

Webb, D.F.: 2000, Understanding CMEs and their source regions. Journal ofAtmospheric and Solar-Terrestrial Physics 62, 1415 – 1426.

Wheatland, M.S., Sturrock, P.A., Roumeliotis, G.: 2000, An Optimization Ap-proach to Reconstructing Force-free Fields. Astrophys. J. 540, 1150 – 1155.doi:10.1086/309355.

Wiegelmann, T.: 2004, Optimization code with weighting function for thereconstruction of coronal magnetic fields. Solar Phys. 219, 87 – 108.doi:10.1023/B:SOLA.0000021799.39465.36.

Wiegelmann, T.: 2007, Computing Nonlinear Force-Free Coronal MagneticFields in Spherical Geometry. Sol. Phys. 240, 227 – 239. doi:10.1007/s11207-006-0266-3.

Wiegelmann, T., Inhester, B., Sakurai, T.: 2006, Preprocessing of Vector Mag-netograph Data for a Nonlinear Force-Free Magnetic Field Reconstruction.Solar Phys. 233, 215 – 232. doi:10.1007/s11207-006-2092-z.

Wiegelmann, T., Thalmann, J.K., Schrijver, C.J., Derosa, M.L., Metcalf, T.R.:2008, Can We Improve the Preprocessing of Photospheric Vector Magne-tograms by the Inclusion of Chromospheric Observations? Solar Phys. 247,249 – 267. doi:10.1007/s11207-008-9130-y.

Wills-Davey, M.J.: 2006, Tracking Large-Scale Propagating Coronal Wave Fronts(EIT Waves) using Automated Methods. Astrophys. J. 645, 757 – 765.doi:10.1086/504144.

Wills-Davey, M.J., Rachmeler, L.A., Sechler, M.: 2008, Observations of anUnusual Propagation in SOHO-EIT. ApJ, submitted.

Xie, H., Ofman, L., Lawrence, G.: 2004, Cone model for halo CMEs: Applicationto space weather forecasting. Journal of Geophysical Research (Space Physics)109(A18), 3109. doi:10.1029/2003JA010226.

Xue, X.H., Wang, C.B., Dou, X.K.: 2005, An ice-cream cone model for coronalmass ejections. Journal of Geophysical Research (Space Physics) 110(A9),8103. doi:10.1029/2004JA010698.



Yang, L., Jin, R., Sukthankar, R., Zheng, B., Mummert, L., Satyanarayanan,M., Chen, M., Jukic, D.: 2007, Learning distance metrics for interac-tive search-assisted diagnosis of mammograms. In: Medical Imaging 2007:Computer-Aided Diagnosis. Edited by Giger, Maryellen L.; Karssemeijer,Nico. Proceedings of the SPIE, Volume 6514, pp. 65141H (2007)., Presentedat the Society of Photo-Optical Instrumentation Engineers (SPIE) Conference6514. doi:10.1117/12.710076.

Yashiro, S., Gopalswamy, N., Michalek, G., St. Cyr, O.C., Plunkett, S.P., Rich,N.B., Howard, R.A.: 2004, A catalog of white light coronal mass ejectionsobserved by the SOHO spacecraft. Journal of Geophysical Research (SpacePhysics) 109(A18), 7105. doi:10.1029/2003JA010282.

Yurchyshyn, V.B., Wang, H., Goode, P.R., Deng, Y.: 2001, Orientation ofthe Magnetic Fields in Interplanetary Flux Ropes and Solar Filaments.Astrophys. J. 563, 381 – 388. doi:10.1086/323778.

Zhao, X.P., Plunkett, S.P., Liu, W.: 2002, Determination of geometri-cal and kinematical properties of halo coronal mass ejections using thecone model. Journal of Geophysical Research (Space Physics) 107, 1223.doi:10.1029/2001JA009143.



Documents

Computer Vision for the Solar Dynamics Observatory · Computer Vision for the Solar Dynamics Observatory ... produce a comprehensive system of automated feature recognition for the