Real-time flare detection in ground-based H α imaging at Kanzelhöhe Observatory
Werner Pötzi, Astrid M. Veronig, Gernot Riegler, Ute Amerstorfer, Thomas Pock, Manuela Temmer, Wolfgang Polanec, Dietmar J. Baumgartner
aa r X i v : . [ a s t r o - ph . S R ] N ov Solar PhysicsDOI: 10.1007/ ••••• - ••• - ••• - •••• - • Real-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory W. P¨otzi · A.M. Veronig , · G. Riegler · U. Amerstorfer · T. Pock · M. Temmer · W. Polanec · D.J. Baumgartner c (cid:13) Springer ••••
Abstract
Kanzelh¨ohe Observatory (KSO) regularly performs high-cadence full-disk imaging of the solar chromosphere in the H α and Ca ii K spectral linesas well as the solar photosphere in white-light. In the frame of ESA’s SpaceSituational Awareness (SSA) programme, a new system for real-time H α dataprovision and automatic flare detection was developed at KSO. The data andevents detected are published in near real-time at ESA’s SSA Space Weather por-tal ( http://swe.ssa.esa.int/web/guest/kso-federated ). In this paper, we describethe H α instrument, the image recognition algorithms developed, the implemen-tation into the KSO H α observing system and present the evaluation resultsof the real-time data provision and flare detection for a period of five months.The H α data provision worked in 99 .
96% of the images, with a mean time lagbetween image recording and online provision of 4 s. Within the given criteria forthe automatic image recognition system (at least three H α images are needed fora positive detection), all flares with an area ≥
50 micro-hemispheres and locatedwithin 60 ◦ of the Sun’s center that occurred during the KSO observing timeswere detected, in total a number of 87 events. The automatically determinedflare importance and brightness classes were correct in ∼ ∼ ◦ . Themedian of the absolute differences for the flare start times and peak times fromthe automatic detections in comparison to the official NOAA (and KSO) visualflare reports were 3 min (1 min). Kanzelh¨ohe Observatory for Solar and EnvironmantalResearch, University of Graz, Austriaemail: [email protected] email: [email protected] email: [email protected] email: [email protected] Institute of Physics/IGAM, University of Graz, AustriaGraz University of Technologyemail: [email protected] email: [email protected] Institute for Computer Graphics and Vision, GrazUniversity of Technology, Austriaemail: [email protected] email: [email protected]
SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 1 ¨otzi et al.
Keywords:
Active regions; Flares, Dynamics; Instrumentation and Data Man-agement
1. Introduction
Solar flares are sudden enhancements of radiation in localized regions on theSun. The radiation enhancements are most prominent at short (EUV, X-rays)and long (radio) wavelengths, with only minor changes in the optical contin-uum emission. However, flares are well observed in strong absorption lines inthe optical part of the spectrum, most prominently in the H α Balmer line ofneutral hydrogen at λ = 656 . et al. , 2011). Flares may or may not occur in association with coronal mass ejec-tions (CMEs). However, the association rate is a strongly increasing function ofthe flare importance, and in the strongest and most geo-effective events typicallyboth occur together (Yashiro et al. , 2006).CMEs, flares and solar energetic particles (SEPs), which are accelerated eitherpromptly by the flare or by the interplanetary shock driven ahead of fast CMEs,are the main sources for severe space weather disturbances at Earth. CMEsare only very limited accessible to observations from ground, due to their faintappearance and the stray light in the Earth atmosphere. They are best trackedin white-light images recorded from coronagraphs on space-based observatories.Flares are regularly observed at X-ray and (E)UV wavelengths from satellites,but they are also well observed from ground-based observatories in the H α spectral line.Besides regular visual detection, reporting and classification of solar H α flaresby a network of observing stations distributed over the globe, and collectionat NOAA’s National Geophysical Data Center (NGDC), there are also recentefforts to develop automatic flare detection routines. The detection methodsrange from comparatively simple image recognition methods based on intensityvariation derived from running difference images (Piazzesi et al. , 2012), region-growing and edge-based techniques (Veronig et al. , 2000; Caballero and Aranda,2013) to more complex algorithms using machine learning (Qahwaji, Ahmed, andColak, 2010, Ahmed et al. , 2013) or support vector machine classifiers (Qu et al. ,2003). These methods have been applied to space-borne image sequences in theEUV and soft X-ray range (e.g, Qahwaji, Ahmed, and Colak, 2010; Bonte et al. ,2013; Caballero and Aranda, 2013), but also to ground-based H α filtergrams(e.g., Veronig et al. , 2000; Henney et al. , 2011; Piazzesi et al. , 2012; Kirk et al. ,2013). The flare classification system used in this paper is based on theH α flare importance classification (ˇSvestka, 1966). Figure 1 shows SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 2 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory Subflare Imp. 1 Imp. 2 Imp. 3 Imp. 4 Ha-Flare Importance Class10 -8 -7 -6 -5 -4 -3 GO E S X -r a y F l ux [ W m - ] C2 C8 M4 M9 X4 BCMX GO E S X -r a y C l a ss Figure 1.
GOES X-ray flares plotted against H α flares observed at KSO. The dark back-ground represents the density of data points, the stars indicate the mean of the logarithmicX-ray flare class the relation between the optical H α flare importance class and theX-ray flare class from the Geostationary Operational EnvironmentalSatellites (GOES). The scatter plot contains all flares observed atKSO during the period 1/1975 - 10/2014 that were located within60 deg from the central meridian. The associated GOES X-ray flareswere automatically identified by the following criteria: the soft X-ray and H α flare peak times are within 10 min and the heliographicpositions are within 10 deg . Figure 1 reveals a high correlation betweenthe H α importance class (defined by the chromospheric flare area; cf.Table 2) and the GOES X-ray class (defined by the peak flux in the1-8 ˚A channel). In total the set comprises 2832 flares with the follow-ing distributions among the classes (H α importance: 81.2% subflares,15.4% importance 1, 2.6% importance 2 and 0.8% importance 3 and 4;GOES X-ray class: 86.0% B and C, 12.5% M and 1.5% X-class flares). Space-based data have the advantage that there are no atmospheric distur-bances (seeing, clouds) degrading the image quality, but there is a delay inthe data availability related to the data downlink. Ground-based data have theadvantage that the data are immediately available for further processing, andcan thus be efficiently used for the real-time detection and alerting of transientevents such as solar flares in the frame of a space weather alerting system -however, with the drawback that the image sequences may suffer from data gaps
SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 3 ¨otzi et al. and bad seeing conditions causing varying image quality. These circumstanceshave to be accounted for by the image recognition algorithms applied.In this paper, we present an automatic image recognition method that wasdeveloped for the real-time detection and classification of solar flares and filamenteruptions in ground-based H α imagery. The algorithms have been implementedinto the H α observing system at Kanzelh¨ohe Observatory (KSO), in order toimmediately process the recorded images and to provide the outcome in al-most real-time. This activity was performed in the frame of the space weathersegment of ESA’s Space Situational Awareness (SSA) programme, and the real-time H α data and detection results are provided online at http://swe.ssa.esa.int/web/guest/kso-federated . In this paper, we concentrate on the automaticflare detection and classification system, which was implemented in the KSOobserving system in June 2013, and present the evaluation of the system for afive month period. The automatic detection of filaments and filament eruptionswill be presented in a subsequent study, as the method is still under improvement(first results are shown in P¨otzi et al. , 2014).The paper is structured as follows. In Sect. 2, we describe the KSO solarinstruments and observations. In Sect. 3, we outline the image recognition al-gorithms developed to automatically identify solar flares in H α images, and tofollow their evolution (in terms of location, size, intensity enhancement andclassification). Sect. 4 outlines how the real-time detection and alerting wasimplemented in the KSO observing system. In Sect. 5, the outcome of the real-time flare detection system is evaluated for a test period of five months from endof June to November 2013. In Sect. 6, we discuss the performance of the system.
2. KSO instrumentation and observations
Kanzelh¨ohe Observatory for Solar and Environmental Research (KSO; http://kso.ac.at ) is operated throughout the year at a mountain ridge in southernAustria near Villach (N 46 ◦ ◦ α spectral line (Otruba and P¨otzi, 2003), the Ca ii K spectral line (Hirtenfellner-Polanec et al. , 2011), and in white-light (Otruba, Freislich, and Hanslmeier,2008). Figure 2 shows an exemplary set of simultaneous KSO imagery in H α ,Ca ii K and white-light recorded on January 6, 2014. All data are publicly avail-able via the online KSO data archive at http://kanzelhohe.uni-graz.at/ (P¨otzi,Hirtenfellner-Polanec, and Temmer, 2013).The observations are carried out by the KSO observing team during 7 daysa week, basically from sunrise to sunset if the local weather conditions permit.All instruments for solar observations are mounted on the KSO surveillancetelescope, which comprises four refractors on a common parallactic mounting(Figure 3). The KSO H α telescope is a refractor with an aperture ratio numberof d/f = 100/2000 and a Lyot band-pass filter centered at the H α spectral line( λ = 656 . SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 4 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory Figure 2.
Solar full-disk observations in white-light (left), Ca ii K (middle) and H α (right)recorded at Kanzelh¨ohe Observatory on January 6, 2014. Figure 3.
Kanzelh¨ohe Observatory (left) and its solar patrol telescope consisting of fourrefractors observing the Sun in H α , Ca ii K and white-light (right). path. The Lyot filter can be tuned by turning the polarizers in narrow boundarieswith little degradation of the filter characteristics. A beam splitter allows theapplication of two detectors at the same time. Currently, the observations aresolely carried out in the center of the H α line.The CCD camera of the H α image acquisition system is a Pulnix TM-4200GEwith 2048 × et al. ,1994; Shine et al. , 1995) to benefit from moments of good seeing. The imagedepth of the CCD camera is 12 bit, which allows observing the quiet Sun andflares simultaneously without overexposing the flare regions. In order to havegood counts statistics under varying seeing conditions and to avoid saturationeffects in strong flares, an automatic exposure control system is in place; theautomatically controlled exposure time lies in the range 2.5 to 25 milliseconds.In the standard observing mode, the observing cadence of the H α telescopesystem is 6 seconds. The plate scale of the full-disk observations is ∼ α images. SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 5 ¨otzi et al.
3. Image recognition algorithm
The developed image recognition algorithms make use of the main characteristicsof the features in single H α images as well as in images sequences. Solar flaresare characterized by a distinct brightness increase of localized areas on the Sun.They reach their maximum extent and maximum intensity typically within someminutes up to some tens of minutes, followed by a gradual decay of the intensitydue to the subsequent cooling of the solar plasma. Flares are categorized inimportance classes based on their total area and their brightness enhancementwith regard to the quiet Sun level.The image recognition algorithm consists of four main building blocks. The preprocessing handles the different intensity distributions, large-scale inhomo-geneities and noise. The feature extraction step defines the characteristic at-tributes of the features to be detected, and how to model them. In the multi-labelsegmentation step, the model is applied to “new”, i.e. previously unseen, imagesin almost real-time. In the postprocessing , every identified object is assignedand tracked via a unique ID, and the characteristic flare parameters are derived(location, area, start/peak time, etc.) In the following we give a basic descriptionof these methods; further details can be found in Riegler et al. (2013) and Riegler(2013).3.1. PreprocessingThe preprocessing has two goals, namely image normalization and feature en-hancement. Across different H α image sequences, the intensity distributions ofthe images are shifted and dilated. These differences in the distributions arisedue to different solar activity levels (e.g. many/few sunspots), seeing conditions,exposure time, etc. As the feature extraction strongly relies on value of the imageintensities, we normalize the image intensities by a zero-mean and whiteningtransformation: µ = 1 | Ω | X x ∈ Ω f ( x ) (1) σ = s | Ω | − X x ∈ Ω ( f ( x ) − µ ) (2) f n ( x ) = f ( x ) − µσ (3)where Ω ⊂ R is the image domain, µ the sample mean and σ the standarddeviation of the input image f , respectively. The normalized image f n is givenby a point-wise subtraction and division by the mean and standard deviation.As a second step in the preprocessing, additive noise and large-scale intensityvariations, caused by the center-to-limb variation and clouds, are removed byapplying a structural bandpass filter. At the core of this particular filteringmethod is the total variation with ℓ fitting term (TV- ℓ ) model (Chan and SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 6 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory Figure 4.
Structural bandpass filter applied to an H α image with clouds. a) Original image f ,b) denoised image v , c) large-scale variations v , d) resulting image u of the structuralbandpass filter. Esedoglu, 2005; Aujol et al. , 2006), which is a signal and image denoising methodbased on minimizing a convex optimization problem given bymin u k∇ u k , + λ k f − u k (4)where f is the noisy observation of the image and u is the minimizer of theoptimization problem. The first term, the total variation norm, regularizes thegeometry of the solution and the second term, the ℓ norm, ensures that thesolution is close to the original image f . Finally, λ is a free parameter that canbe used to control the amount of regularization. The main property of the TV- ℓ model is that it is contrast invariant. In other words, structures from the image aremoved only in terms of their spatial extent and not in terms of their contrast tothe background. To efficiently solve the optimization scheme we use the genericprimal dual algorithm proposed in Chambolle and Pock (2011). SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 7 ¨otzi et al.
It was shown in Chan and Esedoglu (2005) that by solving the TV- ℓ for acertain parameter λ , all structures having a minimal width of λ − are removed inthe regularized image u . We utilize this fact in our structural bandpass filter byfirst removing small-scale noise from the image using a larger λ , which resultsin image v . In the next step, we remove larger structures by again regularizingthe image v using a smaller λ < λ such that the resulting image v s containsonly unwanted large scale structures such as brightness variations, clouds, etc.The final result of the structural bandpass filter is then given by subtractingimage v from image v .Figure 4 illustrates the different steps of the structuralbandpass applied to a sample KSO H α filtergram with clouds.3.2. Feature extractionIn the feature extraction step, two main problems have to be addressed. a) Whatare the characteristic attributes of flares and filaments, i.e. what discriminatesthem from other solar regions? b) How can we efficiently model these attributes?To solve the first problem we assign to each pixel a feature vector. The mostintuitive feature choice is the pixel intensity of the preprocessed images. We uti-lize the fact that filaments appear darker than the background of the H α images,and that sunspots are even darker than filaments and have also different typ-ical geometries compared to filaments (round versus elongated objects). Flaresare defined as objects with distinctly higher intensities than the background.It may also be useful to use the intensities of the pixels within a small localneighborhood. Further, the contrast decreases from the center towardsthe limb. To incorporate this effect, the distance from the solar diskcenter to the pixel location has proven to be useful.
Based on the extracted feature vectors we utilize a Gaussian mixture modelto assign each pixel of an H α image a class probability. For the classes we usethe features “flare”, “filament” and “sunspot”. The remaining part of the imageis summarized in the class “background”.Figure 5 illustrates the class probabilities in a histogram. The data that weuse for the feature extraction and the learning of the model are derived fromlabeled H α images, where an expert annotated a set of KSO H α images byassigning the pixels to the different classes. As one can see from the figure, theprobability distributions of the four classes are not distinctly separated. Theoverlaps between the classes sunspot–filament and background–flare are no se-vere problem, because most of the probability mass is well separable. In contrast,the probability distribution overlap between the classes filament–backgrounddoes cause segmentation problems in the application. Additional methods thatcan be used to arrive at a better distinction of filaments against the backgroundare described in Riegler (2013).3.3. Multi-label segmentationIn principle, each pixel could be assigned to the class with the highest probability,however, this would lead to a very noisy segmentation. In order to regular-ize the final segmentation, we adopt a total variation based multi-label image SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 8 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory Figure 5.
Intensity distribution of the classes sunspot (red), filament (green), background(black) and flare (red). The training examples are derived from preprocessed H α images thatwere annotated by an expert. segmentation model (Chambolle and Pock, 2011):min { u l } Nl =1 N X l =1 Z Ω d |∇ u l | + N X l =1 Z Ω u l q l d x (5)s.t. u l ( x ) ≥ , N X l =1 u l ( x ) = 1 , (6)where the functions u l and q l , l = 1 ...N are the binary class assignment func-tions and class-dependent weighting function, respectively. In the simplest casethe negative logarithm of the class probabilities can be used, but we apply anadditional temporal smoothing by computing an exponential weighted movingaverage over the probabilities.3.4. PostprocessingThe final step of the method is the postprocessing that has two main goals. Thefirst one is the identification of each detected flare (and filament) with a uniqueID, which should remain the same over the image sequence for the very sameobject. The second goal is the derivation of characteristic properties from theidentified objects to categorize them. The identification task is solved by means of a connected-component labelingproblem (Rosenfeld and Pfaltz, 1966). For the tracking of the objects in the H α SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 9 ¨otzi et al.
Figure 6.
Sample KSO H α image from May 10, 2014, in which three different flares aresimultaneously present on the solar disk. The detected flare areas are indicated in differentcolors. Each flare is registered and tracked in time, i.e. from image to image by a unique ID. image sequence, we apply a simple propagation technique. From the segmen-tation we obtain the four binary images u l for the four classes. The next stepis to identify 8-connected pixels that form a group and are separated throughzeros from other groups, and to assign an ID to them. The problem can beefficiently solved with a two-pass algorithm as presented for example in Haralockand Shapiro (1991). In a first pass, temporary labels are assigned and the labelequivalences are stored in a union-find data structure. Then a label equivalenceis detected, whenever two temporary labels are neighbors. In the second pass,the temporary labels are replaced by the actual labels that are given by the rootof the equivalence class. The union-find data structure is a collection of disjointsets and has two important functions. The union function combines two sets,and the find function returns the set that contains a given number. The datastructure can be efficiently implemented with trees.The connected component labeling ensures that every flare (and filament)has a unique ID per image. To guarantee that the ID remains the same throughthe image sequence, we propagate the ID of previous images. Assume that I t ( x ) SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 10 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory Figure 7.
Sample KSO H α image from March 31, 2014 together with the filament detections.Each filament is assigned and tracked by its unique ID (annotated at each filament). is the current component labeled segmentation and { I t − k ( x ) } nk =1 the set of n previous component labeled segmentation results. Then we change the ID of acurrent component I t to the ID j , where j is given by the components of theprevious images that have the most overlap with the component I t . This can beimplemented in a pixel-wise fashion and a simple map data structure. For a givencomponent of the current image and ID I t we iterate all overlapping pixels x ofthe set { I t − k ( x ) } nk =1 . If I t − k ( x ) = 0, we increment the counter for the ID I t − k ( x )in the map. Finally, we assign the ID with the highest counter. Since flares oftenconsist of two or more ribbons, flare detections that are located within a certaindistance (set to 150 arcsec) are grouped to one ID. Figure 6 shows a sample H α image with the flare detections and the assigned flare IDs. A sample H α imagewith the filament detections is shown in Figure 7. SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 11 ¨otzi et al.
4. Implementation at the KSO H α observing system To optimize the process and speed of the real-time data provision and flare detec-tion, different computers are involved that run in parallel, each one performinga specific set of tasks:– camera computer: image acquisition;– workstation 1: quality check, data processing and online data provision;– workstation 2: image recognition, flare detection and alerting.Figure 8 shows a flow diagram of the tasks that are performed on each in-coming H α image, where the columns refer to different machines responsible forcertain tasks. Each H α image is grabbed by the camera computer and sent toworkstation 1, where the image is checked for its quality. If the quality criterionis passed, the image is processed and published on the web server. In parallel,the processed image is also transferred to workstation 2, on which the imagerecognition algorithm is performed. If an event is detected, its characteristicparameters are calculated. In case that the event exceeds a certain threshold(i.e. flare area/importance class), a flare alert is published online at ESA’s SSASWE portal and an alert email is sent out. In the following we give a detaileddescription of the different analysis steps.4.1. Image acquisition, processing and online provision Image grabbing
The image acquisition is done in a fully automated mode,which includes automatic exposure control and the use of the frame selectiontechnique (Shine et al. , 1995). The CCD camera is controlled via a simple userinterface; in standard patrol mode no user interaction throughout the observationday is needed.
Quality check
All H α images grabbed are checked for their quality. Cloudsand bad seeing conditions result in low contrast and unsharp images, whichmay cause difficulties for the image recognition. Since the quality test has tobe performed on each image, a simple algorithm was implemented. The imagequality is measured by three conditions that have to be fulfilled:– The solar disk appears as a sphere with high accurateness: points on thesolar limb are detected by a Sobel edge enhancement filter. A circle isfitted through the detected limb points and the relative error of the radiusis computed.– The large-scale intensity distribution is uniform: the solar image is rebinnedto a 2 x 2 pixels image, and the relative brightness differences of these 4pixels define a measure of the intensity distribution.– The image is sharp: the correlation between the raw image and a smoothedversion of the image is computed. If the raw image is already unsharp, itshows a high correlation with the smoothed image. SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 12 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory Figure 8.
Flow diagram showing the main steps of the data processing pipeline at the KSOH α observing system. The overall process of data acquisition, quality check, processing, imagerecognition, event detection and alerting involves four different machines (camera PC, raidsystem, workstation 1 and 2) and a web server where the results are published. SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 13 ¨otzi et al.
Based on these criteria, the images are classified in three quality groups: good,fair and bad. Only images of quality “good” are sent to the image recognitionpipeline and the online data provision. Images classified as “fair” or “bad” aremoved to a temporary archive and are not considered in the further analysis.However, we note that images of quality “fair” may still be acceptable and usefulfor visual inspection (e.g. for visual flare detection).
Image processing
For all images that remain in the pipeline, decisive param-eters like the disk center, the solar radius, maximum and mean brightnesses, etc.are derived. Together with additional information such as the acquisition time,instrument details, solar ephemeris for the recording time, etc., the images arestored as FITS file (Pence et al. , 2010). These images are on the one hand storedin the KSO data archive, on the other hand they are fed into the subsequentpipeline of real-time data provision and image recognition.
Provision of real-time images and movies on the SSA SWE portal
Each minute an image is selected for the real-time H α display at ESA’s SSASWE portal ( http://swe.ssa.esa.int/web/guest/kso-federated ; a snapshot is shownin Figure 9). The size of the image is reduced to 1024 × α images isgenerated and displayed at the SWE portal.4.2. Image recognition, flare characterisation and alerting Image recognition
Most of the iterative algorithms of the image recognition(cf. Sect. 3) are computationally intensive. However, they can be easily paral-lelized. Thus, it is possible to utilize the computational power of modern graphicprocessing units (GPU). The image recognition algorithm has been implementedin the programming language C ++ and installed on a dedicated machine with ahigh-performance GPU. The system benefits from the large number of processingunits which are used for the highly parallelized computations. In its present form,the algorithm needs about 10 s to process the flare and filament recognition onone 2048 × Event detection and parameter calculation
After the detection of a flareby the image recognition system, its characteristic properties and parametersare derived. For flares, these include the heliographic position, the flare area(which defines the importance class), the brightness class, and the flare startand peak times. These quantities need not only the information of a single H α SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 14 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory Figure 9.
Screenshot of the ESA SSA SWE H α subportal ( http://swe.esa.int/web/guest/kso-federated ). The subportal shows the real-time H α image (middle), a list of the detectedevents of that day (right) and a 360 deg view around the observatory (bottom) for checkingthe observating conditions, and is updated every minute. image but also the information stored in the image recognition log files for theprevious time steps. Handling of simultaneous flares is easily possible as eachflare is identified via a unique ID that is propagated from image to image. InFigure 10 we show a sequence of H α images that were recorded during a 2B classflare that occurred on May 10, 2014 (top panels) together with the segmentedflare regions (bottom panels).The flare area is calculated by the number of segmented pixels with thesame ID. These are subsequently converted by the pixel-to-arcsec scale of thatday to derive the area in millionths of the solar hemisphere, so-called “micro-hemisphere”. The conversion procedure includes the information of the flareposition to correct the effect of foreshortening toward the solar limb. The deter-mined area is then directly converted to the flare importance class (subflares,1, 2, 3, 4) according to the official flare importance definitions (cf. Table 1).For the categorization into the flare brightness classes (Brilliant-Normal-Faint:B-N-F), the intensity values relative to the background are used. To this aim,we compute the mean, standard deviation, maximum and minimum of the pixelintensities within the segmented regions. For each detected feature, we apply anormalisation by the difference between the maximum brightness and the meanbrightness of the feature. SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 15 ¨otzi et al.
Figure 10.
Top: Sequence of H α images showing the evolution of a 2B flare that occurred onMay 10, 2014. Bottom: segmented flare areas. Table 1. H α flare importance classes.H α importance Flare area(micro-hemisphere)S[ubflares] < > To characterize a flare, the evolution of the brightness and the area in each H α image of the sequence has to be analyzed. For illustration, we show in Figure 11the evolution of the area and brightness of a sample 1N flare that occured onOctober 16, 2013. The flare classification is based on the following definitions:1. The flare start is defined as the time when the brightness enhancement isabove the faint flare level for 3 consecutive images.2. The peak time of the flare is defined as the time where the maximum flarebrightness is reached.3. The flare position is defined by the location of the brightest flare pixel at thetime of the flare peak.4. The importance class of the flare is defined via the maximum area of the flare,and is updated when the area exceeds the level of a higher importance class.5. The flare end is defined as the time when the brightness has decreased belowthe faint level for 10 consecutive images or when there is a data gap of morethan 20 minutes. SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 16 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory f l a r e a r ea [ - s o l a r he m i s phe r e s ] Imp.-1 Subflare b r i gh t ne ss [ no r m a li z ed ] FaintNormal S t a r t P ea k Figure 11.
Illustration of the flare parameter calculation for the 1N flare of October 16, 2013.Top: Evolution of the flare area. Middle: Evolution of the flare brightness. The crosses showthe data points, the solid line shows the maximum values for each minute. The thick solid linerepresents times where the intensity is above the faint flare level. The determined start (14:27UT) and peak (14:31 UT) times are indicated by the green and blue bar, respectively. Notethat the maximum brightness and the maximum area do not necessarily occur at the sametime. Bottom: Snapshots of the flare at three different times (indicated by orange vertical lineson the top panels) together with the flare detections.
6. Handling of data gaps: In case of data gaps <
20 min, the flare is consideredto be in an evolving state if the flare brightness after the data gap is higherthan before the gap. Data gaps of >
20 minutes define the end of the flare.
Flare alerts
If a flare is detected that exceeds a certain threshold, i.e. impor-tance class, then a flare alert is published on the ESA SSA SWE portal and an
SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 17 ¨otzi et al. alert email is sent out to registered users. Originally, it was intended to restrictthe flare alerting to events of H α importance class 1 and higher. However, dueto the weak activity cycle 24, we lowered the threshold to subflares exceedinga size of 50 micro-hemispheres, in order to obtain a sufficient statistics for theevaluation. The event list on the ESA SSA SWE portal is updated every time a flare isdetected, when more information on flare becomes available during its evolution(e.g., the peak time) or when a flare that is already listed increases in its im-portance class. The flares are sorted in decreasing start time, so that the mostrecent event appears in the first line of the table. As long as the flare brightnessincreases, no peak time is listed but the event is annotated to be “ongoing” andmarked in red color in the event list in the H α subportal (cf. Fig. 9). When theflare brightness has decreased for >
5. Results
The system of near real-time H α data provision, automatic flare detection andalerting went online on June 26, 2013. In the following we present the resultsfrom evaluating the system for a period of five months from June 26 to November30, 2013, in which it was run with the same set of parameters and definitions.5.1. Real-time data provisionTo validate the online data provision, we evaluated the number of H α imagesthat were recorded by the KSO observing programme and the number of H α images that were provided online in almost real-time at the ESA SWE serviceportal. For this purpose, log files recorded for each image both the observationtime and the time when the image was put online to the SWE service portal.During the evaluation period, we had in total 563 hours of solar observationsat KSO. In total, 395 129 H α images were recorded. 281 806 images (71.3%) wererated as “good”, 20 922 (5.3%) as “fair”, and 92 401 (23.4%) as “bad”. 33 765 H α images (one per minute) out of the “good” sample were provided online at theSWE service portal, whereas 14 were erroneously skipped due to internal datastream errors. This means that in 99 .
96% of the observation time, one image perminute was provided online at the SWE portal. The mean time lag between therecording of an image and its online provision was 3 . ± . In solar cycles 21 to 23 about 10% of H α flares were larger than subflares (Temmer et al. ,2001; Joshi and Pant, 2005). However, this number is much smaller in the current low-activitycycle 24. E.g. in the year 2013, only 35 of a total of 565 flares visually identified at KSO andreported to NOAA were larger than subflares. SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 18 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory α flares, weconsidered all flares that exceeded an area of 50 micro-hemispheres and thatoccurred within 60 ◦ from solar disk center during the KSO observing times. Asdiscussed in Sect. 4.1, only images of quality “good” are fed into the automaticflare detection pipeline. For the automatic detection of a flare, we demand thatit is observed in at least three H α images. Periods of >
20 min containing noimages of quality “good” are defined as data gaps.The data that are needed for the evaluation of the flare detection, classificationand alerting are derived from the log files that are created and updated during theobservations (cf. Fig. 8). The relevant parameters that are derived and evaluatedare:– Flares: heliographic position, start time, peak time, area, importance class,brightness class;– Alerts: time of issue.For the evaluation, we compare the results obtained by the automated imagerecognition system developed, called Surya , against the official flare reportsprovided by the National Geophysical Data Center (NGDC) of the NationalOceanic and Atmospheric Administration (NOAA) and by Kanzelh¨ohe Observa-tory. Both are obtained by visual inspection of the data by experienced observers.The Space Weather Prediction Center (SWPC) of the U.S. Department ofCommerce, NOAA, is one of the national centers for environmental protectionand provides official lists of solar events, online available at . The information on the flare events is col-lected from different observing stations from all over the world. Kanzelh¨ohe Ob-servatory sends monthly flare reports to different institutions, including NGDC/NOAA and the World Data Center (WDC) for Solar Activity (Observatoire deMeudon). The visual KSO flare reports (KSOv) are online available at http://cesar.kso.ac.at/flare data/kh flares query.php . We actually expect that the resultsof the automatic detections are on average closer to the visual KSO flare reportsthan the NOAA reports, as they are based on the data from the same observa-tory. However, it is also important to compare the outcome against the NOAAreports, as they provide an independent set of flare reports.Table 2 in the Appendix lists all flare events (area ≥
50 micro-hemispheres;located within 60 ◦ from disk center) which were detected in quasi real-time bythe automated algorithm during the evaluation period. In total, 87 flares weredetected by Surya; 69 were classified as subflares and 18 as flares of importance1. This list includes 3 false detections (marked in red color in the table), i.e.flares that were detected by Surya but have no corresponding event reported by For flares closer to the solar limb than 60 ◦ from the center of the disk, projection effectsbecome significant in the determination of the flare area. In addition, these flares are mostlikely not relevant for space weather disturbances at Earth. Surya -“the Supreme Light” is the chief solar deity in Hinduism
SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 19 ¨otzi et al.
NOAA or KSOv. In addition, the list in Table 2 includes 7 flares where Suryareported one flare but NOAA and KSOv reported two separate events, as wellas 2 NOAA (4 KSOv) flares, where Surya has split up one flare into two or moreevents.To evaluate the detection ability of Surya, we checked also all NOAA andKSOv flares reported during the KSO observation times (located within 60 ◦ from disk center) that were not detected by Surya. These are in total 60 flares(57 SF, 2 SN and one 1N flare). There are basically two different reasons whythese events were not detected by Surya:1. Data gaps: Less than 3 images of quality “good” were available during theevent, and thus the automatic detection algorithm was not run. 47 flares fallinto this category. We note that single images as well as images of lowerquality may still be sufficient to identify a flare by visual inspection (thoughwith large uncertainties in the derived flare parameters). Indeed, for 19 out ofthese events visual flare reports from KSOv are available (18 subflares, andone flare of importance 1).2. NOAA reports a flare which is not listed by KSOv and where – even aftercareful visual re-inspection of the KSO H α image sequences – we cannotconfirm the appearance of a flare. This applies to 13 events, all of themsubflares.Given the reasons above, these events are not expected to be detected by ourautomatic image recognition system. This means that Surya basically detectedall flares listed by the NOAA and KSOv flare reports, when there were sufficientdata available (i.e. at least three images during an event).The next question we have to address is how accurate are the flare parameterscalculated by the automatic system. The accuracy of the peak flare area deter-mined by Surya cannot be evaluated, since the flare reports do not provide areas.However, the flare area is an intrinsic property which determines the importanceclassification of a flare (cf. Table 1). For the importance classifications we findthat there are 7 cases, in which the importance classes reported by NOAA andKSOv do not conincide. Thus we excluded those events from the evaluationof the importance classes as there is no unique reference value available. Forthe remaining set of flares, we find that in 86% the automatically determinedflare importance class coincides with the class given by NOAA and KSOv. Theincorrectly classified events include 5 flares where Surya obtained a differentimportance class than reported by NOAA and KSOv, and 6 cases where Suryahas split up flares in 2 or more events. For the brightness classification, thereare 15 events where the NOAA and KSOv reports do not coincide. For theremaining set, we find that in 85% Surya determined the correct brightnessclass (cf. Table 2).Figure 12 shows the absolute differences of the heliographic latitude andlongitude of the flare center as obtained by the automatic algorithm againstthe values reported by NOAA and KSOv. The mean of the absolute differencefor the latitude is 1 . ◦ (0 . ◦ ) with respect to the NOAA (KSOv) flare reports,and 1 . ◦ (0 . ◦ ) for the longitude. SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 20 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory Figure 12.
Distribution of the absolute differences of the flare’s heliographic latitude (left)and longitude (right) between the automatic detection by Surya and the NOAA (blue) andKSOv (red) reports.
Figure 13.
Distribution of the absolute differences between start times of the flares detectedby Surya and reported by NOAA (blue) and KSOv (red).
Figure 14.
Distribution of the absolute differences between peak times of the flares detectedby Surya and reported by NOAA (blue) and KSOv (red).
SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 21 ¨otzi et al.
Figure 15.
Distribution of the time differences between the start of the flare and the issue ofthe alert.
Figures 13 and 14 show the distributions of the absolute differences of theflare start times and peak times, respectively, derived by Surya in comparisonto NOAA and to KSOv. For the start and the peak times, the median timedifference is 3 min (1 min) with respect to NOAA (KSO). For 62% (78%) of theflares detected, the derived flare start times lie within ± ± ≥
50 micro-hemisphere, andanother one when it reaches an area of 100 micro-hemispheres, i.e. the thresholdfor an importance 1 flare. In total, we had 14 cases (14%) where erroneously noflare alert emails were issued. These include the 4 flares on July 21, 2013, the 5flares on Oct 11, 2013 and the first 2 flares on Oct 15, 2013, where we had anerror in the automatic email script. In addition, no alert was sent for 3 flaresthat reached exactly the threshold area of 50 micro-hemispheres (Aug 15, 2013,12:03 and and 12:49; Aug 30, 2013, 06:14). The total number of false alerts wassix. Three of the false alerts are related to the false flare detections (indicated inred in Table 2). The other three false alerts were double alerts, i.e. two identicalemails had been sent for one flare.We also evaluated the time between the occurrence of the flare and the issue ofthe alert email. In this respect, occurrence means the time when the flare reaches
SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 22 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe Observatory the threshold area, i.e. an area of 50 micro-hemispheres for a subflare alert, oran area of 100 micro-hemispheres for an importance 1 alert, etc. Figure 15 showsthe distribution of the time difference between the occurrence of the flare andthe issue of the alert email, giving a median of 1 . α image sequences. The detection of a flare demands that a flareis detected in at least three observations. However, if there is a longer data gap,say, between the second and the third image of a sequence, then the alert (whichis issued after the flare detection in the third image) may occur substantiallydelayed with respect to the start of the flare (defined by the time of the firstimage in this sequence).
6. Discussion and Conclusions
The real-time H α data provision worked perfectly fine, with a percentage of99 .
96% provisions online at the ESA SWE portal within less than 4 s of theobservations. The automatic flare detections basically worked in all cases withinthe given criteria, i.e. within the demand that for a positive detection we needat least three H α images of quality “good” during the event. In our 5-monthevaluation period, about 70% of all the H α images observed at KSO (witha regular cadence of 6 s) were classified as “good”. We note that for visualclassifications by an experienced observer also fewer images or images of lowerquality may suffice but result in flare parameters with large uncertainties. This isreflected in the number of 87 flares that were detected by the automatic system,whereas the KSO visual reports included 19 additional events (18 SF, one 1N).The automatically determined flare importance and brightness classificationswere correct in about 85% of the events. The misclassification of 15% is com-parable to the ∼
7% (17%) inconsistencies between the NOAA and KSOv flarereports for the importance (brightness) class. These differences in the officialreports are related to different instruments, seeing conditions and observers thatderived the parameters. The mean of the calculated heliograpic longitude andlatitude of the flare center was consistent with the official flare reports within ∼ ◦ . The median of the absolute differences between the flare start and flarepeak times of the automatic detections in comparison to the NOAA (KSOv)reports were 3 min (1 min). In ∼
90% of the flare alert emails that were sent, thealert was issued within 5 min of the flare start. However, 15% of the expectedalerts had been not sent. The number of false flare detections and alerts was lessthan 6% compared to the total number of (true) alerts issued.We note that our event set consisted mostly of subflares (69) and importance1 flares (18). There was one event reported as 2N by NOAA and KSOv (October10, 2013; cf. Table 2), which was misclassified as 1F by our automatic algorithm.Re-inspection of the processing of this event showed that the area had beencorrectly calculated (i.e. exceeded the threshold to an importance 2 flare), butdue to a large data gap during the observing sequence, the algorithm erroneouslyapplied a wrong time (during the flare decay phase) for the calculation of theimportance class.
SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 23 ¨otzi et al.
We conclude that the automatic flare detection implemented at KSO andprovided online at ESA’s SSA SWE portal provides reliable and near real-timedetection, classification and alerting of solar H α flares. The information on theflare timing, strength and heliographic position (which relates to the magneticconnectivity to Earth) that is derived in near real-time could, e.g., be connectedto SEP models. We note that the largest challenges of the algorithm are actuallythe handling of data gaps, which are the largest source of misclassification of theflare class and the split-up of one flare into more than one. Further systematicevaluation of the system at times of higher solar activity and more frequentoccurrence of larger flares will be valuable in order to test its ability for theautomatic detection of the most severe space-weather effective events. Appendix
Table 2 lists all flares detected by Surya with an area ≥
50 micro-hemispheresand located within 60 ◦ of the solar disk during the period June 26 to November30, 2013 together with the corresponding information from the NOAA and KSOvflare reports. Column 1 gives the observation date, columns 2 − − −
10 the heliographic position, columns 11 −
13 the flare type andcolumn 14 the flare area as determined by Surya. False flare detections by Suryaare marked in red color.
Acknowledgements
This study was developed within the framework of ESA Space Sit-uational Awareness (SSA) Programme (SWE SN IV-2 activity). The authors thank AlexiGlover and Juha-Pekka Luntama for their support, constructive criticism and confidence inthe project.
References
Ahmed, O.W., Qahwaji, R., Colak, T., Higgins, P.A., Gallagher, P.T., Bloomfield, D.S.: 2013,Solar Flare Prediction Using Advanced Feature Extraction, Machine Learning, and FeatureSelection.
Solar Phys. , 157 – 175.
DOI . ADS .Aujol, J.-F., Gilboa, G., Chan, T., Osher, S.: 2006, Structure-texture image decompositionâĂŤ-modeling, algorithms, and parameter selection.
International Journal of Computer Vision (1), 111 – 136.Benz, A.O.: 2008, Flare Observations. Living Reviews in Solar Physics , 1. DOI . ADS .Bonte, K., Berghmans, D., De Groof, A., Steed, K., Poedts, S.: 2013, SoFAST: AutomatedFlare Detection with the PROBA2/SWAP EUV Imager.
Solar Phys. , 185 – 199.
DOI . ADS .Caballero, C., Aranda, M.C.: 2013, Automatic Tracking of Active Regions and Detection ofSolar Flares in Solar EUV Images.
Solar Phys. . DOI . ADS .Chambolle, A., Pock, T.: 2011, A first-order primal-dual algorithm for convex problems withapplications to imaging.
Journal of Mathematical Imaging and Vision (1), 120 – 145.Chan, T.F., Esedoglu, S.: 2005, Aspects of total variation regularized l1 function approxima-tion. SIAM Journal on Applied Mathematics (5), 1817 – 1837.Fletcher, L., Dennis, B.R., Hudson, H.S., Krucker, S., Phillips, K., Veronig, A., Battaglia,M., Bone, L., Caspi, A., Chen, Q., Gallagher, P., Grigis, P.T., Ji, H., Liu, W., Milligan,R.O., Temmer, M.: 2011, An Observational Overview of Solar Flares. Space Sci. Rev. ,19 – 106.
DOI . ADS . SOLA: sniv_solphy_28102014.tex; 2 October 2018; 20:36; p. 24 e a l - t i m e fl a r e d e t ec t i o n i n g r o und - b a s e d H α i m ag i n ga t K a n ze l h ¨o h e O b s e r v a t o r y Table 2. : Automatically detected flares in the period June 26 to November 30, 2013. Flares marked in red color are false detections.
Start time Peak time Position Type AreaSurya NOAA KSOv Surya NOAA KSOv Surya NOAA KSOv Surya NOAA KSOv Surya27/06
SOLA:sniv_solphy_28102014.tex;2October2018;20:36;p.25 ¨o t z i e t a l. Table 3. : Table 2 continued.
Start time Peak time Position Type AreaSurya NOAA KSOv Surya NOAA KSOv Surya NOAA KSOv Surya NOAA KSOv Surya17/08
SOLA:sniv_solphy_28102014.tex;2October2018;20:36;p.26 e a l - t i m e fl a r e d e t ec t i o n i n g r o und - b a s e d H α i m ag i n ga t K a n ze l h ¨o h e O b s e r v a t o r y Table 4. : Table 2 continued.
Start time Peak time Position Type AreaSurya NOAA KSOv Surya NOAA KSOv Surya NOAA KSOv Surya NOAA KSOv Surya15/10
SOLA:sniv_solphy_28102014.tex;2October2018;20:36;p.27 ¨o t z i e t a l. Table 5. : Table 2 continued.
Start time Peak time Position Type AreaSurya NOAA KSOv Surya NOAA KSOv Surya NOAA KSOv Surya NOAA KSOv Surya27/10
SOLA:sniv_solphy_28102014.tex;2October2018;20:36;p.28 eal-time flare detection in ground-based H α imaging at Kanzelh¨ohe ObservatoryHaralock, R.M., Shapiro, L.G.: 1991, Computer and robot vision , Addison-Wesley, ???.Henney, C.J., MacKenzie, D., Hill, F., Mills, B., Pietrzak, J.: 2011, Solar Flare Detection WithSWIFT and Real-time GONG H-alpha Images. In:
AAS/Solar Physics Division Abstracts , 2233.
ADS .Hirtenfellner-Polanec, W., Temmer, M., P¨otzi, W., Freislich, H., Veronig, A.M., Hanslmeier,A.: 2011, Implementation of a Calcium telescope at Kanzelh¨ohe Observatory (KSO).
CentralEuropean Astrophysical Bulletin , 205 – 214. ADS .Joshi, B., Pant, P.: 2005, Distribution of H α flares during solar cycle 23. Astron. Astrophys. , 359 – 363.
DOI . ADS .Kirk, M.S., Balasubramaniam, K.S., Jackiewicz, J., McNamara, B.J., McAteer, R.T.J.: 2013,An Automated Algorithm to Distinguish and Characterize Solar Flares and AssociatedSequential Chromospheric Brightenings.
Solar Phys. , 97 – 111.
DOI . ADS .Otruba, W., P¨otzi, W.: 2003, The new high-speed H α imaging system at Kanzelh¨ohe SolarObservatory. Hvar Observatory Bulletin , 189 – 195. ADS .Otruba, W., Freislich, H., Hanslmeier, A.: 2008, Kanzelh¨ohe Photosphere Telescope (KPT).
Central European Astrophysical Bulletin , 1 – 8. ADS .Pence, W.D., Chiappetti, L., Page, C.G., Shaw, R.A., Stobie, E.: 2010, Definition of the FlexibleImage Transport System (FITS), version 3.0.
Astron. Astrophys. , A42.
DOI . ADS .Piazzesi, R., Berrilli, F., Del Moro, D., Egidi, A.: 2012, Algorithm for real time flare detection.
Memorie della Societa Astronomica Italiana Supplementi , 109. ADS .P¨otzi, W., Hirtenfellner-Polanec, W., Temmer, M.: 2013, The Kanzelh¨ohe Online Data Archive.
Central European Astrophysical Bulletin , 655 – 660. ADS .P¨otzi, W., Riegler, G., Veronig, A., Pock, T., M¨ostl, U.: 2014, A system for near real-time detec-tion of filament eruptions at Kanzelh¨ohe Observatory. In:
IAU Symposium , IAU Symposium , 519 – 520.
DOI . ADS .Priest, E.R., Forbes, T.G.: 2002, The magnetic nature of solar flares.
Astron. Astrophys. Rev. , 313 – 377. DOI . ADS .Qahwaji, R., Ahmed, O., Colak, T.: 2010, Automated Feature Detection and Solar FlarePrediction Using SDO Data. In: , COSPAR Meeting , 2877. ADS .Qu, M., Shih, F.Y., Jing, J., Wang, H.: 2003, Automatic Solar Flare Detection Using MLP,RBF, and SVM.
Solar Phys. , 157 – 172.
DOI . ADS .Riegler, G.E.: 2013, Flare and Filament Detection in Halpha Solar Image Sequences. Master’sthesis, Graz University of Technology.Riegler, G., Pock, T., P¨otzi, W., Veronig, A.: 2013, Filament and Flare Detection in H α imagesequences. ArXiv e-prints . ADS .Roggemann, M.C., Welsh, B.M., Ford, S.D., Stoudt, C.A.: 1994, Frame selection for adaptiveoptics imaging through atmospheric turbulence. In: Schulz, T.J., Snyder, D.L. (eds.)
Im-age Reconstruction and Restoration , Society of Photo-Optical Instrumentation Engineers(SPIE) Conference Series , 42 – 53.
ADS .Rosenfeld, A., Pfaltz, J.L.: 1966, Sequential operations in digital picture processing.
Journalof the ACM (4), 471 – 494.Sammis, I., Tang, F., Zirin, H.: 2000, The Dependence of Large Flare Occurrence on theMagnetic Structure of Sunspots. Astrophys. J. , 583 – 587.
DOI . ADS .Shine, R.A., Tarbell, T., Title, A., Scharmer, G., Simon, G., Brandt, P., Berger, T.: 1995,Frame Selection Techniques for Solar Movies. In:
AAS/Solar Physics Division Meeting , Bulletin of the American Astronomical Society , 957. ADS .Temmer, M., Veronig, A., Hanslmeier, A., Otruba, W., Messerotti, M.: 2001, Statistical analysisof solar H α flares. Astron. Astrophys. , 1049 – 1061.
DOI . ADS .ˇSvestka, Z.: 1966, Optical Observations of Solar Flares.
Space Sci. Rev. , 388 – 418. DOI . ADS .Veronig, A., Steinegger, M., Otruba, W., Hanslmeier, A., Messerotti, M., Temmer, M., Brunner,G., Gonzi, S.: 2000, Automatic Image Segmentation and Feature Detection in Solar Full-Disk Images. In: Wilson, A. (ed.)
The Solar Cycle and Terrestrial Climate, Solar and Spaceweather , ESA Special Publication , 455.
ADS .Wiegelmann, T., Thalmann, J.K., Solanki, S.K.: 2014, The Magnetic Field in the SolarAtmosphere.
ArXiv e-prints . ADS .Yashiro, S., Akiyama, S., Gopalswamy, N., Howard, R.A.: 2006, Different Power-Law Indices inthe Frequency Distributions of Flares with and without Coronal Mass Ejections.
Astrophys.J. Lett. , L143 – L146.
DOI . ADS ..