Solar Magnetic Feature Detection and Tracking for Space Weather Monitoring
Paul A. Higgins, Peter T. Gallagher, R.T. James McAteer, D. Shaun Bloomfield
aa r X i v : . [ a s t r o - ph . S R ] J un Solar Magnetic Feature Detection and Tracking forSpace Weather Monitoring
P.A. Higgins ∗ , P.T. Gallagher ∗ , R.T.J. McAteer, D.S. Bloomfield Astrophysics Research Group, School of Physics, Trinity College Dublin, Dublin 2,Ireland
Abstract
We present an automated system for detecting, tracking, and catalogingemerging active regions throughout their evolution and decay using
SOHO
Michelson Doppler Interferometer (MDI) magnetograms. The
SolarMoni-tor Active Region Tracking (SMART) algorithm relies on consecutive imagedifferencing to remove both quiet-Sun and transient magnetic features, andregion-growing techniques to group flux concentrations into classifiable fea-tures. We determine magnetic properties such as region size, total flux, fluximbalance, flux emergence rate, Schrijver’s R -value, R ∗ (a modified version of R ), and Falconer’s measurement of non-potentiality. A persistence algorithmis used to associate developed active regions with emerging flux regions inprevious measurements, and to track regions beyond the limb through mul-tiple solar rotations. We find that the total number and area of magneticregions on disk vary with the sunspot cycle. While sunspot numbers are aproxy to the solar magnetic field, SMART offers a direct diagnostic of the sur-face magnetic field and its variation over timescale of hours to years. SMARTwill form the basis of the active region extraction and tracking algorithm forthe Heliophysics Integrated Observatory (HELIO). Keywords: active regions, feature detection, region growing algorithm,space weather ∗ Corresponding Authors
Email address: [email protected] (P.A. Higgins)
Preprint submitted to Advances in Space Research November 5, 2018 . Introduction
The automatic identification and characterization of solar features is ofgreat importance to both solar activity monitoring and space weather opera-tions. This has become a particular issue due to the high spatial and temporalresolution solar imagers, such as those flown on the
Project for On-Board Au-tonomy 2 ( PROBA2 ) and
Solar Dynamics Observatory ( SDO ), which willforce data providers to distribute subsets of their science products instead ofthe full image data set. Traditionally, solar feature catalogs were created byhand, using visual recognition to record the position, size, and other prop-erties of features (e.g., Carrington, 1854). An early attempt to overcomethis was SolarMonitor (Gallagher et al., 2002), which labels active regions(ARs) in solar images using National Oceanic and Atmospheric Administra-tion (NOAA) numbers and locations cataloged by the NOAA Space WeatherPrediction Center. More recently, researchers have begun to catalog fea-tures using automated methods. The European Grid of Solar Observations (EGSO; Bentley et al., 2002), for example, catalogs solar features using H α and Ca II K images and a neural network algorithm (Zharkova et al., 2005;Zharkova & Schetinin, 2005).One of the first applications of automated image processing techniques toAR identification is the Automated Region Selection Extraction algorithm(McAteer et al., 2005a). This algorithm creates a binary mask of featuresusing a static noise threshold applied to a line-of-sight (LOS) magnetogram.A sub-image is extracted, centered on the pixel with the highest value. Closedcontours enclosing an area centered on the seed are grouped as a region. Thedetected region is saved and removed from the magnetogram. The pixel ofthe next highest value is selected and the process repeated. Some savedregions are associated with NOAA cataloged regions which may be trackedacross the disk. More recently, LaBonte et al. (2007) extract ARs using full-disk magnetograms that are smoothed by roughly one supergranule diameter.Region candidates are tested for bipolar flux and east-west orientation. Adynamic noise threshold is calculated using the median of average magneticfield values for a series of annuli centered on the AR candidate. The ARboundary is chosen by comparing the average magnetic field values of smallerannuli with the calculated noise threshold. Using annuli to test for region (HELIO). SMARTcombines extraction techniques with AR magnetic property determinations(Section 2), region tracking, and cataloging (Section 3). A cross comparisonof SMART and NOAA detections as well as a discussion of errors in propertymeasurements and feature tracking test cases are presented in Section 4. Our DIMag
Segment FeatureStructCharac-terize ClassifyCatalogTrack
VSummary Flow Chart
SecondaryInput /OutputInputOperationFinalOutputExtractRegion
Figure 1: Flow chart summarizing the SMART algorithm processing method. conclusions and prospects for future work are then given in Section 5.
2. Feature Extraction
The SMART method of operation is summarized in Figure 1. Initially,magnetograms are segmented into individual feature masks (Section 2.1). Acharacterization algorithm is then run on each extracted region to determinefeature properties (Section 2.2). These region properties are subsequentlyused to classify the form of solar features (Section 2.3). The final output is aset of data structures for each magnetogram, including each feature present.The following sub-sections provide details on the operations outlined above.
The segmentation process depicted in Figure 2 begins with two consecu-tive
Solar and Heliospheric Observatory ( SOHO )/Michelson Doppler Inter-ferometer (MDI; Scherrer et al., 1995) full-disk, line-of-sight (LOS), level 1.8magnetograms. Nominally these are 96 minutes apart, but there are sporadicgaps in the MDI data set (only rarely is there an entire day with no data).4
DIB t-Δt
S T LProcessDiffRotateto t Binary
Segmentation Flow Chart V Grow M t-Δt SubtractTransDiffMasks
Grow M t MDIB t S T LProcessBinary IndexedGrownMask(IGM)
GrowIndex VV V V
V V V
Grow
S T LProcessGaussian SmoothNoise ThresholdLine of Sight Correct
Index
Assign Unique Integersto Contiguous FeaturesDilate Binary MaskDiffMasksLocate Transient FeaturesRemove Transient FeaturesSubtractTransDiffMasksNon-zero pixels are set to one.BinaryIGM t,i
Figure 2: Flow chart summarizing the magnetogram segmentation method.
We use two magnetograms recorded close in time to remove transient fea-tures and extract time-dependent properties. The magnetogram of interest(Figure 3 A) is denoted as B t and the previous magnetogram as B t − ∆ t . If∆ t is greater than one day, the detections are discarded.Magnetograms are first checked for problems using properties extractedfrom the Flexible Image Transport System (FITS) data file headers, suchas the spacecraft roll angle and the number of missing pixel values. Mag-netograms are rotated as necessary, so that solar north points up, usingnearest neighbor sampling interpolation, while those with missing values arediscarded. A solar energetic particle (SEP) event which occurs during amagnetogram exposure results in many bright pixels scattered about the im-age. This does not interfere with the AR detection, as the bright pixels aresmoothed out, but can affect magnetic property determinations.5 igure 3: Processing steps for an example feature extraction on 25 November 2003. A)Calibrated megnetogram B t clipped to ± B t with gaussian smoothing and noise thresholding. C) Mask ( M f,t ) with transient filteringand area threshold of 50 pixels. D) Final indexed grown feature mask, IGM t,i . We first apply smoothing, a noise threshold, and a LOS correction, re-spectively, to the data (Figure 3 B). This set of operations is represented by
STL Process in Figure 2. The smoothing operation is necessary to removeephemeral regions that have size scales on the order of 10 Mm (Hagenar,2001), which corresponds to 7 MDI pixels at disk center. To this end, B t − ∆ t and B t are convolved with a 10 ×
10 pixel kernel containing a 2D gaussianwith a full-width at half-maximum (FWHM) of 5 pixels.We use a static threshold to remove the background. Figure 4 showsthe variation in the monthly averages of maximum values of quiet-Sun (QS)magnetic field recorded throughout cycle 23. The maximum value varies byroughly 5 G over the cycle which is less than the monthly standard devia-tion of these maxima, so a static threshold is acceptable. The mean of themaximum unsigned QS magnetic field values is ∼
70 G. Figure 5 shows asmoothed AR and nearby QS region contoured at ±
70 G. The histogramshows the distributions of magnetic field values for the AR and QS regions,6 M ax o f Q S M ag F i e l d [ G ] S I DC o f Sun s p o t s Figure 4: The maximum of quiet-Sun magnetic field values over solar cycle 23. Each pointis the mean maximum value for a month of magnetograms (nominally two per day, butless for particularly active periods). The error bars are the standard deviations of eachmonth’s set of values. The continuous gray line is the smoothed, monthly sunspot numberfrom Solar Influences Data Analysis Center (SIDC; http://sidc.oma.be). including the difference between the two distributions. Thresholding at the ±
70 G level removes small features which have been smoothed out by thegaussian convolution but maintains extended strong-field features, such asbipolar and plage regions. Pixels in B t − ∆ t and B t with absolute values lessthan 70 G are zeroed.In the case where magnetic fields are primarily vertical to the solar sur-face, the LOS component of the field is reduced toward the limb. As such,a feature with the same magnetic field strength and orientation with respectto the solar surface will appear lower in magnitude when located towardthe solar limb than at disk center. This LOS effect is corrected at eachMDI pixel using a cosine correction factor (McAteer et al., 2005a). Afterthis stage, B t − ∆ t data is differentially rotated to time t to correct for featuremotions due to solar rotation using the latitudinal dependence derived inHoward et al. (1990).The corrected magnetograms are made binary by setting all pixels withmagnetic field values above the ±
70 G threshold equal to one, yielding masks M t − ∆ t and M t . Features consisting of less than 50 pixels and those which arenot present in both masks are removed by the following operations (Figure 37 igure 5: A comparison of magnetic field value distributions for a quiet Sun and solarfeature region. Top left : Magnetogram of NOAA 8086, gaussian smoothed using a FWHMof 5 pixels.
Top right : A nearby region of quiet Sun in the same full-disk image.
Bottom :The feature and quiet-Sun unsigned magnetic field distributions. The thick red line is thedifference between the quiet-Sun and AR distributions and the vertical dash-dotted linedenotes 70 G.
C). Firstly, each mask is dilated by 10 pixels to allow for region expansion.Secondly, the binary masks are subtracted such that non-zero pixels in thedifference mask identify features only occurring in M t − ∆ t or M t . These tran-sient features are subsequently removed from the un-grown version of M t ,which is then dilated by 10 pixels to form M f,t (Figure 3 D). Individual con-tiguous features in M f,t are indexed by assigning ascending integer values(beginning with one) in order of decreasing feature size. The segmentationoutput is an indexed grown mask ( IGM t ), as shown by the thick red box inFigure 2. 8ata Type Identifier ExplanationFeature Array B t,i extracted feature magnetogram IGM t,i extracted feature mask HG t heliographic position map A cos,t,i IGM t,i cos( HG t × (1.4 Mm / pixel) − Φ t,i B t × A cos,t,id Φ dt | t,i ( | B t |−| B t − ∆ t | ) × A cos,t,i ∆ t Property Value HG pos,t,i P pix ( B t,i × HG t ) P pix ( HG t × IGM t,i ) B max,t,i maximum value of B t,i B min,t,i minimum value of B t,i B tot,t,i P pix B t,i B tot uns,t,i P pix | B t,i | µ, σ , γ, κ mean, variance, skewness, kurtosis A tot,t,i P pix A cos,t,i Φ + ,t,i P pix (Φ t,i > − ,t,i P pix (Φ t,i < uns,t,i P pix | Φ t,i | Φ imb,t,i | (Φ + ,t,i −| Φ − ,t,i | ) | Φ uns,t,i d Φ dt | net,t,i P pix d Φ dt | t,i Table 1: Feature magnetic properties derived from characterization processing.
The aim of SMART is to characterize ARs in a manner which does notmake theoretical assumptions or require many observations of the same fea-ture. Our design is adaptable, so that the software may produce initial re-sults in near-realtime for operational purposes, but allows the retrospectiveaddition of complex property measurements (e.g., magnetic helicity). Theserequirements define criteria for the selection of initial property calculations.There are many AR properties that may be derived from magnetograms. Asubset of these are derived from 96 minute LOS data and those output bySMART are included in Tables 1 and 2.The SMART characterization process utilizes the feature mask retrievedby the methods outlined in the previous section, following the proceduredetailed in Figure 6. The property measurements are derived from the mag-9
DIB t-Δt
T LProcessDiffRotateto t Characterization Flow Chart B t,i Multiply Cos AreaWeightMDIB t T LProcessSubtract ÷ d t B max,t,i , B min,t,i ,B tot,t,i , B tot uns,t,i ,Statistical Moments ExtractExtract V V V IGM t,i B t d B / d ttimage MultiplyAcos,t, i Multiply A totExtractExtract Φ +,t,i, Φ -,t,i , Φ uns,t,i , Φ imb,t,i, d Φ / d t t,i, Statistical Moments
V V VV
VVV Φ t,iimage Extract V B t,i d Φ / d t it V VV V V ExtractT LProcess Noise ThresholdLine of Sight Correct i denotes indexing of IGM.Non-region pixels are zeroed. V V Figure 6: Flow chart summarizing the feature magnetic property characterization method. netogram taken at time, t which is processed in the manner detailed below.We subscript the mask containing all features by i to extract a single featuremask, IGM t,i . A cosine-weighted area map, A cos,t,i is derived which correctspixel areas to solar surface area rather than plane-of-sky area, and is summedto yield total feature area, A tot,t,i .Full-disk magnetograms B t − ∆ t and B t are processed as in Section 2.1(thresholding, LOS correction, B t − ∆ t differentially rotated to time t ), butwithout smoothing. Single features are extracted for magnetic property de-termination using the indexed grown mask, B t,i = IGM t,i × B t , (1)10ata Type Identifier ExplanationFeature Array M P SL,t,i polarity separation line mask M P SL,thin,t,i thinned polarity separation line maskProperty Value L P SL,t,i P pix M P SL,thin,t,i L sg,t,i L P SL,t,i >
50 G Mm − R t,i R -value R ∗ t,i P pix ( M P SL,t,i ∗ Gauss D ) × B t,i W L sg,t,i non-potentiality gauge W L ∗ sg,t,i P pix M P SL,t,i × ∇ B t,i Schrijver (2007) Falconer et al. (2008)Table 2: Feature magnetic properties derived from polarity separation line characteriza-tion. yielding an array where all pixels but those in the feature are set to zero. Theprocessed magnetograms are subtracted and divided by their time separationto yield a map of the temporal change in field strength, dB/dt | t , leading upto time t . This is combined with A cos,t,i to determine the flux emergencerate, d Φ /dt | t,i , of feature i .The extracted B t,i is used to extract other properties from feature i (asdetailed in Table 1) such as statistical moments of the magnetic field and theminimum and maximum magnetic field values ( B min,t,i and B max,t,i ). B t,i ismultiplied by A cos,t,i to derive the total positive, negative, and unsigned flux(Φ + ,t,i , Φ − ,t,i , and Φ uns,t,i ), the relative flux imbalance (Φ imb,t,i ), and the netflux emergence rate ( d Φ /dt | net,t,i ).The extracted feature magnetogram, B t,i , is also used to derive proper-ties based on the polarity separation line (PSL). Figure 7 summarizes theextraction of feature properties related to PSLs and Table 2 lists the proper-ties derived. Initially, the feature is segmented into its positive and negativecomponents. These components are used to create a positive and negativemask, each of which is dilated by 4 pixels. The two masks are summed andthe region of mask overlap becomes the PSL binary mask, M P SL,t,i . The al-gorithm then thins M P SL,t,i to one pixel ( M P SL,thin,t,i ) and sums the non-zeropixels to determine the PSL length ( L P SL,t,i ). L sg,t,i is obtained by summingonly those pixels which have ∇ B t,i >
50 G Mm − , where ∇ B t,i is calculatedby numerical differentiation using 3-point Lagrangian interpolation. We also11 t,i TrinaryGrow
Polarity Separation Line Characterization Flow ChartM
PSLt,iSum GaussSmooth ≥ 2
Binary B Δ Thin V V M tri,t,i + ThreshWLsg R-Maskt,iLPSL ExtractExtract ≥ Binary Values below threshold are set to 0 those aboveare set to 1. M tri,t,i| - |Grow M PSL,thint,i ThreshR-value B Δ Multiply ≥ ThreshBinary Δ MultiplyExtractWLsg,t,i MultiplyRt,iMultiplyR*t,iExtract
V VVV
V V V
V VVV VVV V V Extract B Δ ExtractLSGThresh Δ WLsgt,i * Multiply ≥ ThreshBinary Δ Figure 7: Flow chart summarizing the quantities derived from feature polarity separationlines. calculate the R -value ( R t,i ) as presented in Schrijver (2007) and the W L sg gauge ( W L sg,t,i ) as presented in Falconer et al. (2008), both of which usespecific gradient and magnetic field thresholding when extracting the PSL.Finally, using M P SL,t,i we calculate R ∗ t,i which is a more sensitive version ofthe R -value, since it contains no gradient thresholding and the magnetic fieldthreshold of ±
70 G is much lower than the ±
150 G used in Schrijver (2007).The algorithm convolves M P SL,t,i with a 20 ×
20 pixel kernel containing a 2Dgaussian with a FWHM of 10 pixels, which is multiplied by B t,i and summedto achieve R ∗ t,i . Similarly, an alternative of W L sg,t,i , W L ∗ sg,t,i is calculatedby applying the Falconer et al. (2008) method, but using a magnetic fieldthreshold of ±
70 G and no gradient threshold.12 eatureCandi-datePolarityBalanceUni-polar AmountFlux d Φ / d t V Classification Flow Chart
Multi-polar
MultiSmallDecay UniLargeEmerge
VV VV V (+)(-) LessThanGreater VV UniSmallDecayMultiLargeDecayUniLargeDecay MultiSmallEmergeUniSmallEmergeMultiLargeEmerge
Figure 8: Flow chart summarizing the feature classification method.
A set of data structures is created for each magnetogram including theabove mentioned properties of each extracted feature (used for classification;Section 2.3) and the feature’s heliographic location and time of measurement(used for feature tracking; Section 3).
At this stage the SMART algorithm has characterized the properties ofeach automatically extracted feature. The classification process uses theseproperties to discriminate between various feature types, which are saved inthe algorithm output. Extracted features are initially grouped (as shown inFigure 8) into two catagories: features with a flux imbalance greater than90% are classified unipolar (U), while those having less than 90% are classi-fied multipolar (M). After polarity balance, the total unsigned magnetic flux(Φ uns,t,i ) is tested. Features with Φ uns,t,i greater than 10 Mx are classified aslarge (L), while features with Φ uns,t,i less than 10 Mx are classified as small(S). Finally, the sign of d Φ dt | t,i is tested to determine if features are increasingin flux (emerging, E) or decreasing in flux (decaying, D). The classification13cheme results in eight possible feature classifications which are then alsoattributed to common magnetic feature designations: MLE and MLD aredenoted evolving ARs; MSE and USE are denoted emerging flux concentra-tion (EF); MSD, USD, and are denoted decaying flux concentrations (DF),and finally, ULE and ULD are denoted plage (PL). These common designa-tions are also saved in the algorithm output, allowing one to make a quickassessment of which regions on disk are interesting from a monitoring pointof view. For example, EFs may become ARs and evolving ARs may produceactivity during their evolution, while PL and DF are not likely to produceactivity.
3. Tracking
Having detected various solar features, the SMART algorithm associatesfeatures across different time intervals. Spatial and temporal information isused to track features between consecutive images (Section 3.1) and aroundthe far-side of the disk between consecutive solar rotations (Sections 3.2).Features are then cataloged using the time of their first detection and theirclassification (Section 3.3).
The set of features in a magnetogram is compared with the previous fivemagnetogram sets to associate previously catalogued features with the cur-rent set. Feature positions ( HG pos,t,i ) are differentially rotated, using thelatitudinal dependence derived in Howard et al. (1990), to the same time t and features matched when their heliographic separations are less than 5 de-grees. Features having one classification in previous sets may be associatedwith features having a different one in the current set. Thus, the SMARTalgorithm is capable of tracking possible ARs (MLE, MLD) back to theirfirst emergence as an EF (MSE). Decaying features may also be associatedwith features previously denoted as possible ARs, allowing ARs to be fol-lowed through their final stages of evolution. Fragmentation often occurs inthese late stages which SMART allows for since it does not preclude multiplefeatures from being associated with a single previous feature. If one featuresplits into two, each resulting fragment will be associated with the originalfeature if the resulting fragment positions are within the matching thresholdof the original. A letter is appended to the catalog name of each additionalassociated feature so that individual fragments may be differentiated.14 .2. Far-side Passage Features are tracked beyond the limb through multiple solar rotations tostudy their evolution from emergence to decay. We calculate the rotationperiod, P rot,i , for each feature at time t , which depends on its heliographiclatitude due to differential solar rotation. The feature position is comparedto those in the five magnetogram sets centered on time ( t + t ) − P rot,i usingthe method in the previous section. In this approach the feature positionis essentially rotated to a longitude of +70 degrees then back one full solarrotation, where t is the time taken for the feature to rotate to 70 degreesheliographic longitude from its position at t . In this way the feature isconstrained to have been previously detected just before west limb passage,which increases the efficiency of the algorithm. There are two identifications recorded for each detected feature in a mag-netogram at time, t . The first, i is obtained from IGM t and denotes thetwo-digit size order of the feature. A feature within a single magnetogramis uniquely identified by i . The second identification is the static catalogname, YYYYMMDD.MG.NN, where YYYY is the four digit year, MM isthe two digit month, and DD is the two digit day. The next two charactersspecify the feature type: MG denotes a photospheric magnetic feature. Thisscheme can be expanded to incorporate coronal holes (CH), filaments (FI),and transient features such as flares (FL) and coronal mass ejections (CE) inEUV images. Finally, NN is i when the feature is given a static catalog name.This catalog name is determined once for each feature upon first detection,and is used for all measurements of the same feature as it is tracked throughtime.
4. Results and Discussion
Figure 9 summarizes a comparison of NOAA and SMART AR detectionsover the cycle 23, including numbers of detections and total feature area ondisk. The top panel shows the total number of regions detected in each dataset, arranged in monthly bins; the correlation coefficient between the (un-binned) daily data is 0 .
88. We estimate the frequency of divergence betweenthe detections using the ratio of NOAA to SMART AR daily detections:the ratio is between zero and one 6%, equal to one 22%, between one andtwo 60%, and greater than two 12% of the time. We see a smaller number15 R e g i o n s / M o n t h S M AR T A re a / M o n t h [ M m ] N O AA A re a / M o n t h [ M m ] NOAASMART
Figure 9: A comparison of NOAA and SMART AR detections (binned by 1 month) overcycle 23. The data gap in 1998 is due a the loss of communications with the SOHOspacecraft for several months. of SMART than NOAA AR detections 72% of the time; the mean ratio ofNOAA to SMART AR detections is 1 .
5. This is likely due to the joining oftwo or more nearby sunspot groups by SMART, while NOAA identifies eachindividual sunspot group, regardless of proximity . As such, SMART detec-tions are representative of isolated magnetic systems, while NOAA detectionsrepresent a feature recognition approach. Additionally, NOAA records de-tections by eye, and only if they are visible in intensity data (i.e., if there isa magnetic flux concentration with no sunspot SMART may detect a regionwhen NOAA does not). The bottom panel shows the total area of NOAAregions scaled to the total area of SMART regions. In fact, the NOAA area NOAA may also detect very weak sunspots which may have a Φ uns,t,i too small fordesignation as an AR by SMART.
16s lower by a factor of ∼
50, since only the low-intensity area of sunspots issummed, while the area of extended magnetic features is recorded in SMARTdetections. Number and area are the only two feature properties which canbe directly compared, as NOAA data do not contain any magnetic propertymeasurements.The determination of the magnetic properties of a feature is affectedby MDI magnetogram noise levels, calibration, strong field saturation, andLOS effects. The feature detection itself is generally not affected by thesephenomena, however. The instrument noise threshold of MDI is nominally ±
20 G (Scherrer et al., 1995). This is smoothed by the gaussian convolution,and the segmentation threshold of ±
70 G is well above this. For magneticproperty calculations, a gaussian convolution is not used, so noise contributes20 G to the uncertainty of pixel values above the QS threshold of 70 G. ForSMART region 20031026.MG.11 observed at disk center on 25 November2003, which is found to have a A tot,t,i of 3 . × Mm and a Φ uns,t,i of5 . × Mx, the uncertainty is 7 . × Mx, or 5%.Some calibration issues with the MDI data used by SMART are discussedin Wang et al. (2009). It was found that the 2008 calibration of level 1.8data has been partially corrected, in that it does not suffer from a diskcenter-to-limb variation like the 2007 calibration. However, MDI may largelyunderestimate the magnetic field as the ratio of MDI values to those retrievedfrom
Hinode /Solar Optical Telescope data was found to be ∼ .
7. This doesnot affect feature detections since the effect is consistent throughout the dataset, but could contribute a considerable error of ∼
30% for any magnetic fieldor flux measurements.Strong magnetic field saturation in MDI data is discussed in Liu et al.(2007). It is estimated that this phenomenon occurs in ∼
5% of ARs, inwhich the magnetic field measurements in the umbral areas of very strongsunspots behave non-linearly. In extreme cases, the umbra may appear tohave a smaller magnetic field than the surrounding penumbra. In reality, thefield should continue to increase in the umbra, but in level 1.8 data showingNOAA 9002 at disk center, saturation is clearly observed at ∼ . × Mm progressing to the edge of the solar disk. The LOS area is measured17 igure 10: Tracking of 20031026.MG.11 as it rotates around the Sun from 26 October to26 December 2003. at longitude increments of 3 degrees and modified by the SMART cosine areacorrection. This is compared to the disk center area of the spot, resulting inan over correction of ∼
3% when the centroid reaches 60 degrees longitude.18 igure 11: Feature detection and tracking cases which diverge from NOAA. A) Two bipolarregions join and subsequently fragment. B) Several small bipolar regions merge into anAR complex. C) A bipolar region is first detected as two unipolar features and then as asingle bipolar region.
This error depends on morphology and will be more acute for complex featureboundaries. The over correction increases quickly to ∼
40% as the feature istracked toward the limb.An example of the SMART method of feature tracking and cataloging19s shown in Figure 10. Region 20031026.MG.11 is tracked from 26 October2003 to 26 December 2003. The AR rotates beyond the west limb and isdetected again upon returning at the east limb twice. Although the AR istracked to subsequent solar rotations its catalog name remains the same whenit returns. NOAA first detects this AR on 28 October 2003 designating it asNOAA 10488. When the region returns it is designated a new region num-ber, NOAA 10507 and is renamed upon the second return as NOAA 10525.SMART’s persistent naming through multiple rotations allows independentmeasurements of the same feature to be grouped into a single time plot.The top panels in Figure 10 show MDI magnetograms of the region(clipped at ± uns,t,i ),heliographic longitude ( HG pos,t,i ), PSL length ( L P SL,t,i ), and R value ( R t,i )extracted from 20031026.MG.11. Vertical dotted green (blue) lines denotecrossings at ±
60 degrees of the leading (trailing) edge of the feature; in thesecond time plot, the green (blue) curve tracks this leading (trailing) edge intime. In the plot of PSL length, the black curve sums the length of all de-tected PSL segments ( L P SL,t,i ), while the light-blue curve sums those havinga gradient above 50 G Mm − ( L sg,t,i ). Finally, the plot of R-value shows R ∗ t,i in black and R t,i in blue.The stability of the algorithm is estimated using the plot of Φ uns,t,i be-tween days 25.6 (20 November 14:24 UT) and 33.7 (28 November 16:48 UT).A quadratic fit is subtracted to remove the long timescale variation, resultingin an array of residuals. The two-sigma error of the residuals is determinedto be 2 . × Mx or 3% around the mean of Φ uns,t,i . The stability estimateis particular to this example, as cases such as those shown in Figure 11 couldresult in much larger short timescale variation.There are several recurrent SMART feature tracking cases which divergefrom what would be expected of NOAA (Figure 11). The SMART trackingalgorithm allows features to converge and split apart. However, there may beside-effects, such as when a fragment separates from a larger feature and isgiven a new catalog name, due to the centroids of the two being greater thanthe tracking association threshold (top row). Also, an active region complexmay be detected when there are multiple strong field ARs in close proximity(middle row). Finally, a bipolar region which is significantly disjointed andweak may not be properly grouped into a single region (bottom row). Here20e see an example where each polarity is detected as a separate region. Asthis work is designed to aid in flare forecasting, many examples of each ofthese cases may be studied to determine if they possess unexpected flaringproperties. Also, their evolution maybe studied by tracking the featuresfrom first emergence. The frequency of occurrence for these special casescan be estimated using the data and analysis of Figure 9: when N NOAA isgreater than N SMART
SMART is likely grouping regions into AR complexes(or identifying NOAA ARs as EF or DF), and when N SMART is greater than N NOAA
SMART may be detecting individual unipolar features when NOAAgroups them into bipolar regions.
5. Conclusions
The SMART algorithm allows one to monitor ARs on the solar diskin near-realtime and perform extensive studies on AR magnetic properties.SMART is unique among automated AR extraction algorithms in that itallows the temporal analysis of magnetic properties from birth and throughmultiple solar rotations. Future work will include the analysis of trends inAR evolution over the solar cycle. This is a largely untouched subject thatbegs important questions, such as whether ARs are born destined to flare orrandomly evolve to become flare-active. This may also provide new insightsinto the behavior of the solar dynamo.Previous algorithms include some of the functions performed by the SMARTalgorithm, such as feature and magnetic parameter extraction. However,new utilities are incorporated into the SMART code, such as day-to-day andmultiple rotation feature tracking. Extensive AR properties such as area( A tot,t,i ) and total magnetic flux (Φ uns,t,i ) are determined, as are intensiveproperties such as the maximum magnetic field ( B max,t,i ) and statistical mo-ments ( µ, σ , γ, κ ). Some algorithms, including LaBonte et al. (2007) onlydetect the largest regions, while others like Colak & Qahwaji (2009) only de-tect ARs with sunspots in white-light images. All current algorithms trackARs using visually identified NOAA specifications. The SMART algorithmis independent from these specifications and needs no human interventionto detect and track ARs. Additionally, it utilizes an improved feature cata-loging system which incorporates the date of first detection and the featuretype.The SMART algorithm will be used to create a comprehensive catalogof features present in magnetograms covering the entirety of solar cycle 2321nd will be adapted to use SDO /Helioseismic and Magnetic Imager data. Apipeline version of the algorithm will output detections for inclusion in theHeliophysics Event Knowledgebase . Additionally, it will form part of HE-LIO. In this application, ARs tracked using SMART will be associated witha chain of features and events propagating throughout the heliosphere, suchas EUV loops, flares, CMEs, magnetic disturbances and storms detectablein Earth’s aurorae and ground-based magnetometer data, as well as distantparticle instruments such those on the Voyager and Mercury Surface, SpaceEnvironment, Geochemistry, and Ranging ( MESSENGER ) spacecraft.The magnetic properties of ARs retrieved by the SMART algorithm willalso be used for flare forecasting. While the magnetic complexity of ARs isknown to be an important predictor of flare activity (Sammis et al., 2000;Schrijver, 2007; McAteer et al., 2005b; Conlon et al., 2008), recent work byWelsch et al. (2009) shows that extensive magnetic properties outperformintensive properties as predictors of AR flare activity. One of SolarMonitor’scurrent flare-forecasting algorithms assumes Poisson statistics (Moon et al.,2001; Wheatland, 2001; Gallagher et al., 2002) and relies on historical flaringrates from 1988 to 1996 for each McIntosh sunspot classification (McIntosh,1990). This will be superseded by a statistical forecasting algorithm thatmakes use of extensive AR magnetic properties determined by SMART.Any forecasting algorithm which makes use of magnetic properties outputby SMART will need to take into account several sources of error. Randomerrors including magnetogram noise and algorithm stability for the examplepresented in Section 4 result in an error of ±
5% and ±
3% in Φ uns,t,i , respec-tively. This will not affect the forecasting potential of properties involvingΦ uns,t,i for a sufficiently large sample of regions. Calibration errors in MDIresult in an underestimate of the true magnetic field on average by ∼ . Acknowledgements This research is supported by ESA/PRODEX and a grant from the ECFramework Programme 7 (HELIO). RTJMcA (FP6) and DSB (FP7) areMarie Curie Fellows at TCD. We would like to thank the SOHO team formaking both their data and analysis software publicly available and to ac-knowledge the participants of the first ‘Forecasting the All-Clear’ meeting(April 22-24, 2009) who provided helpful comments and insights upon thepresentation of this work. We would also like to show our appreciation tothe two anonymous referees whose comments helped to improve this paper.
References
Bentley, R.D., & EGSO Consortium, EGSO - The European Grid of SolarObservations, In: Ed. A. Wilson (ed.) Solar variability: From Core toOuter Frontiers. Proc. 10th European Solar Physics Meeting, Noordwijk:ESA Publications Division, SP-506, 923-926, 2002.Carrington, R.C., On the Distribution of the Solar Spots in Latitudes sincethe Beginning of the Year 1854, with a Map, Monthly Notices of the RoyalAstronomical Society, 19, 1-3, 1854.Colak, T., & Qahwaji, R., Automated Solar Activity Prediction: A hybridcomputer platform using machine learning and solar imaging for auto-mated prediction of solar flares, Space Weather, 7, 6001-+, 2009.Colak, T., & Qahwaji, R., Automated McIntosh-Based classification ofsunspot groups using MDI images, Solar Physics, 248(2), 277-296, 2008.Conlon, P.A., Gallagher, P.T., McAteer, R.T.J., Ireland, J., Young, C.A.,Kestener, P., Hewett, R.J., & Maguiree, K., Multifractal Properties ofEvolving Active Regions, Solar Physics, 248, 297-309, 2008.Conlon, P.A., McAteer, R.T.J., Gallagher, & Fennell, J.F., MultifractalProperties of Evolving Active Regions, Astrophysical Journal, submitted,2008.Falconer, D. A., Moore, R. L., & Gary, G. A., Magnetogram Measures ofTotal Nonpotentiality for Prediction of Solar Coronal Mass Ejections fromActive Regions of Any Degree of Magnetic Complexity, Astrophysical Jour-nal, 689, 1433-1442, 2008. 23allagher, P.T., Moon, Y.J., & Wang, H., Active-Region Monitoring andFlare Forecasting I.