Issue 
J. Space Weather Space Clim.
Volume 6, 2016
Statistical Challenges in Solar Information Processing



Article Number  A2  
Number of page(s)  13  
DOI  https://doi.org/10.1051/swsc/2015044  
Published online  25 January 2016 
Research Article
Image patch analysis of sunspots and active regions
I. Intrinsic dimension and correlation analysis
^{1}
Electrical Engineering and Computer Science Department, University of Michigan, Ann Arbor, MI 48109, USA
^{2}
SIDC, Royal Observatory of Belgium, 1180
Brussels, Belgium
^{3}
National Solar Observatory, Boulder, CO 80303, USA
^{*} Corresponding author: krmoon@umich.edu
Received:
9
March
2015
Accepted:
13
December
2015
Context. The flare productivity of an active region is observed to be related to its spatial complexity. Mount Wilson or McIntosh sunspot classifications measure such complexity but in a categorical way, and may therefore not use all the information present in the observations. Moreover, such categorical schemes hinder a systematic study of an active region’s evolution for example.
Aims. We propose finescale quantitative descriptors for an active region’s complexity and relate them to the Mount Wilson classification. We analyze the local correlation structure within continuum and magnetogram data, as well as the crosscorrelation between continuum and magnetogram data.
Methods. We compute the intrinsic dimension, partial correlation, and canonical correlation analysis (CCA) of image patches of continuum and magnetogram active region images taken from the SOHOMDI instrument. We use masks of sunspots derived from continuum as well as larger masks of magnetic active regions derived from magnetogram to analyze separately the core part of an active region from its surrounding part.
Results. We find relationships between the complexity of an active region as measured by its Mount Wilson classification and the intrinsic dimension of its image patches. Partial correlation patterns exhibit approximately a thirdorder Markov structure. CCA reveals different patterns of correlation between continuum and magnetogram within the sunspots and in the region surrounding the sunspots.
Conclusions. Intrinsic dimension has the potential to distinguish simple from complex active regions. These results also pave the way for patchbased dictionary learning with a view toward automatic clustering of active regions.
Key words: Sun / Active region / Sunspot / Data analysis / Classification / Image patches / Intrinsic dimension / Partial correlation / CCA
© K.R. Moon et al., Published by EDP Sciences 2016
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
1. Introduction
Active regions (ARs) in the solar atmosphere have intense and intricate magnetic fields that emerge from subsurface layers to form loops which extend into the corona. When active regions undergo external forcing such as flux emergence and rearrangement, the system may destabilize. The stored magnetic energy is then suddenly released as accelerated particles (electrons, protons, ions) and an increase in radiation called a flare is observed across the entire electromagnetic spectrum (Phillips 1991).
The morphology of sunspots is correlated with flare occurrence and has therefore received a lot of attention. The Mount Wilson classification scheme (Hale et al. 1919) groups sunspots into four main classes based on the magnetic structure, that is, on the relative locations and sizes of concentrations of opposite polarity magnetic flux. The sunspots with simplest morphology belong to the unipolar α and the bipolar β groups. More complex morphologies are described as βγ when a bipolar sunspot is such that a single northsouth polarity inversion line cannot divide the two polarities. When a βγ sunspot group contains in addition a δ spot, that is, umbrae of different polarities inside a single penumbra, it is labeled as a βγδ group. The presence of a δ configuration, where large values of opposite polarity exist close together, was identified as a warning of the buildup of magnetic energy stress with an increased probability of a large flare (Mayfield & Lawrence 1985; Sammis et al. 2000). McIntosh (1990) proposes another classification scheme containing 60 classes, thus describing the magnetic structure in greater details. The McIntosh classification is the basis for several flare forecasting methods which estimate the flare occurrence rate from historical records of flares and active region classes (Bornmann & Shaw 1994) possibly combining such information with observed waiting time distribution between flares (Gallagher et al. 2002; Bloomfield et al. 2012).
The McIntosh and Mount Wilson classifications are in general carried out manually, and this results in inconsistencies that stem from human observation bias as well as nonreproducible catalogs. To overcome these caveats, some supervised machine learning methods have been proposed to automatically classify sunspot groups according to these schemes. Stenning et al. (2013) extract various measurements from continuum and magnetogram images, and then feed these into a machine learning classifier which reproduces the Mount Wilson classification. Colak & Qahwaji (2008) employ neural networks and supervised classification techniques to reproduce the McIntosh scheme and use those results in a flare forecasting system (Colak & Qahwaji 2009). While these approaches reduce the human bias, they do not use the information present in sunspot images in an optimal way and make the study of AR dynamic behavior impractical.
Several attempts were made to find quantitative descriptors of an active region’s complexity. McAteer et al. (2005) showed that fractal dimension of an active region alone cannot distinguish between the various Mount Wilson classes. The generalization to multifractal spectrum, where each scale has its own fractal dimension, allowed to study in greater details the evolution of active region in view of distinguishing between quiet and flareproductive active regions. Box counting methods (Abramenko 2005; Georgoulis 2005; Conlon et al. 2008) as well as more accurate methods based on continuous wavelet transform (Conlon et al. 2010; Kestener et al. 2010) were employed. Continuous wavelet transforms and energy spectrum were also used with a similar purpose in Hewett et al. (2008), McAteer et al. (2010).
Wavelet basis functions act as a microscope to describe local discontinuities and gradients in an image, and Ireland et al. (2008) used two multiresolution analyses to compute at various length scales the gradients of the magnetic field along lines separating opposite polarities. Using a data set of about 10 000 magnetogram images, they showed that, at all length scales, those gradients increase going from α to β, βγ, and βγδ classes.
However, a wavelet analysis is known to generate artifacts due to the particular shape of the specific wavelet functions. Signal representations based on a set of redundant functions called a dictionary were therefore introduced (Mallat & Zhang 1993). Elad & Aharon (2006) proposed the use of a small sized dictionary to find a sparse representation of patches. Specifically, a patch is a m × mpixel neighborhood, and a patch analysis of a npixel image will process the m^{2} × n data matrix that collects the overlapping patches. See Figure 1 for a representation.
Fig. 1. An example patch from the edge of a sunspot in a continuum image and its column representation. 
As an example of image patch analysis, Elad & Aharon (2006) considered the problem of denoising an image corrupted by additive Gaussian noise. They computed a sparse representation of patches over a dictionary, thus effectively denoising the patches. The dictionary itself may either be fixed a priori or learned from the corrupted patches. An estimate of the noisefree image is then obtained by averaging the denoised overlapping patches. Elad & Aharon (2006) showed that dictionary learning methods based on patch analysis are more flexible and provide superior results in the context of image denoising.
In this paper, we carry out a patch analysis of a set of sunspots and active region magnetogram images that span the four main Mount Wilson classes. We estimate the intrinsic dimension of the local patches and show how it relates to the Mount Wilson classification. We also study patterns of local correlation using partial correlation and canonical correlation analysis, which reveal some characteristics of simple and more complex active regions. Such analysis also serves as a preparation to an unsupervised clustering of active region using patchbased dictionary learning, which will be presented in a companion paper (Moon et al. 2015).
Section 2 describes our data set. Unlike previous works, our approach combines information from two modalities: photospheric continuum images and magnetograms, both obtained by the Michelson Doppler Imager (MDI) on board the Solar and Heliospheric Observatory (SOHO) spacecraft. We consider 424 active regions spanning the four main Mount Wilson classes. We use SMART masks (Higgins et al. 2011) to delineate the boundaries of magnetic active regions, and the STARA algorithm (Watson et al. 2011) which provides masks for umbrae and penumbrae from the continuum images. These two masks enable us to differentiate between pixels belonging to the actual sunspots and pixels featuring the region surrounding the sunspots.
In Section 3, the intrinsic dimension of the image patches extracted from the two modalities is estimated using both linear and nonlinear methods. The linear method relies on Principal Component Analysis (PCA; Jolliffe 2002), while the nonlinear method relies on a kNearestNeighbor graph approach (Costa & Hero III 2006; Carter et al. 2010). The latter method also estimates the local intrinsic dimension, which has several advantages over a global estimate. We show that the intrinsic dimension is related to the complexity of the sunspot groups.
Section 4 identifies the spatial and modal interactions of the patches at different scales by estimating the partial correlation and by using canonical correlation analysis (CCA; Muller 1982; Nimon et al. 2010). This gives insight into relationships that may exist between active region complexity and the correlation patterns.
This paper expands and refines some of the work in Moon et al. (2014). Whereas Moon et al. (2014) used fixed size square pixel regions centered on the sunspot group as input to the analyses, in this paper SMART detection masks are used. A larger set of images is considered in all methods, which enables us to analyze the relationships of intrinsic dimension and correlation with AR complexity. We also explore the partial correlation of patches, which was not included in Moon et al. (2014).
2. Data
The data used in this study are taken from the Michelson Doppler Imager (MDI) instrument (Scherrer et al. 1995) on board the SOHO Spacecraft.
Within the time range of 1996–2010, we select a set of 424 ARs as follows. Using the information from the Solar Region Summary reports compiled by the Space Weather Prediction Center of NOAA http://www.swpc.noaa.gov/ftpdir/forecasts/SRS/, we consider ARs located within 30° of the solar meridian. We looked at a maximum of 200 instances per Mount Wilson types α, β, βγ, and βγδ. Out of this first selection, we removed AR with a longitudinal extent smaller than four degrees, and finally we checked if MDI continuum and magnetogram data were available. This provides us with a number of ARs in each Mount Wilson class as given by Table 1. In our analysis, we also divide the ARs into two groups: simple ARs (α and β) and complex ARs (βγ and βγδ).
Number of each AR per Mount Wilson class. Simple ARs include α and β groups while complex ARs are βγ and βγδ groups.
ARs are observed using two modalities: photospheric continuum images and magnetogram. SOHOMDI provides almost continuous observations of the Sun in the whitelight continuum, in the vicinity of the Ni I 676.78 nm photospheric absorption line. These photospheric intensity images are primarily used for sunspot observations. MDI data are available in several processed “levels”. We use level1.8 images and rotate them with North up. SOHO provides two to four MDI photospheric intensity images per day with continuous coverage since 1995. We also use the level1.8 lineofsight (LOS) MDI magnetograms, recorded with a nominal cadence of 96 min. The magnetograms show the magnetic fields of the solar photosphere, with negative (represented as black) and positive (as white) areas indicating opposite LOS magneticfield orientations.
As stated in Section 1, SMART masks (Higgins et al. 2011) are used to determine the boundaries of magnetic active regions from MDI magnetograms. Those masks are applied also on continuum images to determine the surrounding part of the sunspot that is affected by magnetic fragments as seen in magnetogram images. Similarly, the STARA algorithm (Watson et al. 2011) provides masks for sunspots (umbrae and penumbrae) from MDI continuum and those masks are applied on magnetogram images to determine the AR cores corresponding to the sunspots. Combining these two types of masks provides thus two sets of pixels within each AR: those belonging to the sunspots themselves as found by STARA and those belonging to the magnetic fragments (or background) within an AR as found by the difference set between the SMART and STARA masks.
As in Moon et al. (2014) we use image patch features to account for spatial dependencies using square patches of pixels. Thus if a SMART mask of an image has n pixels and we use a m × m patch, the corresponding continuum data matrix X is m^{2} × n where the ith column contains the pixels in the patch centered at the ith pixel. The magnetogram data matrix Y is formed in the same way and the full data matrix is with size 2m^{2} × n. We let z_{i} denote the ith column of Z. The images from both modalities are also normalized prior to analyzing them.
In image patch analysis, the size of the patch should be no larger than the smallest feature that is to be captured. Otherwise, the relevant feature may be suppressed. Additionally, large patches lead to highdimensional estimates which suffer in accuracy from “the curse of dimensionality”, which refers to the fact that the number of observations must increase at least linearly in the number of parameters for accurate estimates to be possible in statistical inference (Bühlmann & Van De Geer 2011). Since some sunspot and active region features can be quite small and to limit the effects of high dimensionality on our analysis, we primarily use 3 × 3 patches in each modality although larger patches are used in Section 4 when analyzing spatial correlations in the images.
3. Intrinsic dimension estimation
The goal of this section is to determine the number of intrinsic parameters or degrees of freedom required to describe the spatial and modal dependencies using image patches. We consider 3 × 3 patches within both the continuum and magnetogram images giving an extrinsic dimension of 18. The intrinsic dimension will determine how redundant these 18 dimensions are. In addition, intrinsic dimension provides an indicator of complexity which we compare against the Mount Wilson classification, similarly to what McAteer et al. (2005) and Ireland et al. (2008) did using fractal dimension and gradient strength along polarity separating lines, respectively. More details on the concept of intrinsic dimension on manifolds are included in Appendix A.1.
It is also important to know whether linear analyses can be accurately applied to the data or whether nonlinear techniques are required. Linear methods have been applied successfully to solar images before such as in Dudok de Wit et al. (2013). However, it is not guaranteed that natural images are best represented using linear methods as there are cases where nonlinear models have superior performance (Dobigeon et al. 2014). Thus this is important to investigate both for further analysis of the data as in Moon et al. (2015) and for the correlation analysis in Section 4. If the data lie on a nonlinear subspace and we perform a linear analysis of the data (e.g. partial correlation, canonical correlation analysis, or principal component analysis), then the results will be only a linear approximation of the true relationships and dependencies of the data. Nonlinear methods of analysis would be necessary to obtain higher accuracy in this case. To answer this question, we estimate the local intrinsic dimension using a method appropriate for linear subspaces and a method appropriate for any (linear or nonlinear) smooth subspace and then compare the results.
3.1. PCA: a linear estimator
Principal Component Analysis (PCA) (Jolliffe 2002) finds a set of linearly uncorrelated vectors (principal components) that can be used to represent the data. PCA has been used previously for various purposes in solarphysics and spaceweather literature, e.g. to study the background and sunspot magnetic fields (Lawrence et al. 2004; Cadavid et al. 2008; Zharkova et al. 2012), for analysis of solar wind data (Holappa et al. 2014), or to reduce dimensionality (Dudok DeWit & Auchère 2007).
In PCA, the principal components are the eigenvectors of the covariance matrix Σ:where x and y are random vectors of dimension 9, x being a patch from the continuum image, and y the corresponding patch from the magnetogram image. The eigenvalues indicate the amount of variance accounted for by the corresponding principal component. A linear estimate of intrinsic dimension is the number of principal components that are required to explain a certain percentage of the variance.
By nature, PCA is a global operation and so it provides a global estimate of the intrinsic dimension. We can obtain more local estimates by performing PCA separately on the areas within the sunspots and on the magnetic fragments. These areas are separated using the STARA and SMART masks.
3.2. kNN: a general estimator
The general method we use is a kNearestNeighbor (kNN) graph approach with neighborhood smoothing (Costa & Hero III 2006; Carter et al. 2010). The intuition behind the method is that we grow the kNN graph from a point z_{i} by adding an edge from z_{i} to z_{j} if z_{j} is within the k nearest neighbors of z_{i}. The growth rate of the total edge length of the graph is related to the intrinsic dimension of the data in a way that enables us to estimate it.
One advantage of the kNN method, in contrast to global methods such as Levina & Bickel (2004), is it provides an estimate of the local intrinsic dimension by limiting the growth of the graph to a smaller neighborhood. This provides an estimate of intrinsic dimension at each pixel location in the image which allows us to more easily visualize the intrinsic dimension estimates. Additionally, when the number of samples within a region of interest is small (such as within a small sunspot), this local method provides more accurate estimates of intrinsic dimension than applying a global method (such as PCA) since the inclusion of the neighboring pixels results in a higher number of samples. Technical explanation of the kNN method and more details on local intrinsic dimension are given in Appendix A.1.
3.3. General results
We estimate the intrinsic dimension of the image patches within the sunspots and magnetic fragments for all 424 ARs using both the kNN approach and PCA, where the extrinsic dimension of the joint patches is 18. Figure 2 shows two examples of the estimated local intrinsic dimension using the kNN method and the corresponding continuum and magnetogram images. One set of images corresponds to an α group while the other set is a βγδ group. In these examples, areas with more spatial structure, such as within the sunspots, have lower intrinsic dimension. Fewer parameters are required to accurately represent structured data than noise and so the intrinsic dimension is lower.
Fig. 2. Examples of the estimated local intrinsic dimension using the kNN method for an α group (top) and a βγδ group (bottom). Regions with more spatial structure have lower intrinsic dimension. 
Table 2 provides the mean and standard deviation of the intrinsic dimension estimates within the sunspots and magnetic fragments. These statistics are also provided for ARs within the main Mount Wilson classes (α, β, βγ, and βγδ). We provide PCA results for the cases where we estimate the intrinsic dimension as the number of components required to explain 97% and 98% of the variance, respectively. For the kNN method, we provide the results in two ways. For one, we take the mean of local intrinsic dimensions within each image (separating the “sunspot” from the “magnetic fragments”) and then calculate the mean and standard deviation of these means. The statistics in this category correspond to the mean and standard deviation of the average intrinsic dimension of each image and are more directly comparable to the PCA results. However, these results may be affected slightly by small sunspot groups. For the other approach, we pool all of the local estimates (again separating sunspots from magnetic fragments) and then calculate the mean and standard deviation. These results correspond to the mean and standard deviation of the pixels within each region and category and are less affected by small sunspot groups.
Estimated intrinsic dimension results for different groups of ARs in the form of mean ± standard deviation. The complex ARs have higher intrinsic dimension within the sunspots than the simple ARs but lower intrinsic dimension within the magnetic fragments.
From Table 2, it is clear that the intrinsic dimension is lower within the sunspots than in the magnetic fragments for all methods. This is expected as there is more spatial structure within the images inside the sunspots than in the magnetic fragments, especially in the continuum image.
The average PCA estimate with a 97% threshold and the average mean kNN estimate give similar results inside the sunspot while the average 98% PCA estimate is closest to the average mean kNN estimate within the magnetic fragments. If linear methods were not sufficient to represent the spatial and modal dependencies, we would expect the PCA results to be much higher than the kNN results when using comparable thresholds as more linear than nonlinear components would be required to accurately represent the data. However, this close agreement between the general and linear results suggests that linear methods are sufficient and that linear dictionary methods would be appropriate for these data.
3.4. Patterns within the Mount Wilson groups
For both the PCA and kNN methods, the average estimated intrinsic dimension is lower within the sunspots in α groups than in the more complex groups such as βγδ. This is consistent with Figure 2 and may be related to the lower complexity of α groups. These exhibit more spatially coherent images, which can be described using a lower number of basis elements, and hence have a lower intrinsic dimension.
Within the magnetic fragments, the opposite trend occurs where the less complex groups have higher intrinsic dimension. This suggests that the magnetic fragments are fewer, weaker, and less structured outside of the α and β groups compared to the more complex regions, leading to a more noiselike background in their magnetic fragments. This hypothesis is supported by the normalized histograms of the mean kNN estimates of intrinsic dimension and the normalized histograms of the pooled kNN estimates in Figures 3 and 4. The histograms of mean intrinsic dimension show that within the magnetic fragments, α groups generally have higher mean intrinsic dimension than βγδ groups. In fact, no α groups have a mean intrinsic dimension less than 7.5 within the magnetic fragments. However, the normalized histograms of the individual patch estimates show a significant number of patches with intrinsic dimension less than 7.5 within the fragments. This suggests that for each α group, the majority of the patches have higher intrinsic dimension in the magnetic fragments and are thus more noiselike. In contrast, there are some βγδ groups where the mean intrinsic dimension of the magnetic fragments is lower (<7.5) and so these magnetic fragments are dominated by patches with more structure.
Fig. 3. Normalized histograms of mean estimated intrinsic dimension of α, β, βγ, and βγδ groups using the kNN method. The distributions of intrinsic dimension differ by complexity with simpler AR groups having higher (resp. lower) intrinsic dimension within the sunspot (resp. magnetic fragments). 
Fig. 4. Normalized histograms of pooled local estimates of intrinsic dimension of α, β, βγ, and βγδ groups using the kNN method. The distributions of intrinsic dimension differ by complexity with simpler AR groups having higher (resp. lower) intrinsic dimension within the sunspot (resp. magnetic fragments). 
Table 2 also shows that the standard deviation of the estimates within the sunspots decreases as the complexity increases as measured by the Mount Wilson classification scheme. The histograms in Figures 3 and 4 can be used to determine the cause. From the histograms, it is clear that within the sunspots the intrinsic dimension of α groups does not have a Gaussian distribution. In this case, most of the estimates are between 3 and 5. However, there are a significant number of outliers with intrinsic dimension greater than 5. The presence of these outliers contributes to the high standard deviation. This is in contrast to the intrinsic dimension of βγ and βγδ groups inside the sunspot which have fewer outliers and thus smaller standard deviations.
The outliers in the α groups correspond to small sunspots. The number of pixels within the α sunspots with average intrinsic dimension ≥6 ranges between 10 and 53 with a median of 16. In these cases, the spatial structure of the sunspots may be more similar to the magnetic fragments than the spatial structure of larger sunspots. Thus the intrinsic dimension is higher in the small sunspots.
A similar phenomenon occurs within the β groups. Note that in Table 2, the average and standard deviation of the mean intrinsic dimension of the β groups within the sunspots is higher than for all other groups. This is also caused by a few outliers that have high average intrinsic dimension due to the small size of the sunspots. When individual local intrinsic dimension estimates of the patches from these small sunspots are pooled with the estimates from all other β patches, the average intrinsic dimension is more aligned with that of the other Mount Wilson types. Additionally, ignoring the biggest outliers in the mean intrinsic dimension (defined as having mean intrinsic dimension >6.25) gives an average mean intrinsic dimension of 4.6 for the β groups which is more aligned with the other groups.
The distribution of intrinsic dimension within the magnetic fragments also differs by complexity based on Figures 3 and 4. The complex ARs have more patches and images with lower intrinsic dimension than the simple sunspots which is consistent with Table 2.
In summary, based on the estimated intrinsic dimension of the image patches, relatively few parameters are required to accurately represent the data. We have found that the distribution of local intrinsic dimension varies based on the complexity of the sunspot group with the more complex sunspots having higher (resp. lower) intrinsic dimension within the sunspot (resp. magnetic fragments). Additionally, the standard deviation of the intrinsic dimension is higher within the sunspot in the simpler sunspots than the complex ARs. This is due to the presence of small sunspots among the simpler ARs that tend to have less spatial structure and thus a higher intrinsic dimension than typical sunspots. We have also shown that linear methods should be sufficient to accurately analyze the data.
4. Spatial and modal correlations
The results in the previous section indicate that linear methods are likely sufficient to represent the spatial and modal dependencies within a sunspot. We therefore analyze the linear correlation over patches using partial correlation and canonical correlation analysis (CCA).
The partial correlation is proportional to the inverse of the correlation matrix and analyzes the pixeltopixel correlation when the influence of all other pixels has been removed. It provides insight into how large a patch should be used to sufficiently capture the spatial and modal correlations in future analysis.
Canonical correlation analysis on the other hand is determined by finding the most correlated linear combinations of pixels from each image, solved as a generalized eigenvalue problem, which is useful for determining the degree of mutual correlation between two modalities. If the two modalities are independent, there is no benefit in processing them together, while if the two modalities are strongly dependent, processing only one of the modalities is sufficient since the other modality would not contain any additional information.
4.1. Partial correlation: methodology
The partial correlation measures the correlation between two random variables while conditioning on the remaining random variables. The intuition behind partial correlation can be best explained with the linear regression concept. Suppose you want to compute the partial correlation between two variables X_{1} and X_{2} given a set of variables . First, compute the linear regression using variables in to explain X_{1} and obtain the associated residuals . Proceed similarly for X_{2} and get residuals . The partial correlation between X_{1} and X_{2} is then equal to the (usual) correlation between and , for which the effect of variables has been removed.
In our context, let x be a patch from the continuum image and y be the magnitude (entrywise absolute value) of the corresponding patch from the magnetogram. The partial correlation matrix and its offdiagonal elements can be derived from the inverse correlation matrix (see Appendix A.2). We use the magnitude of the magnetogram data since both positive and negative polarities affect the continuum image in similar ways.
4.2. Partial correlation: results
Figure 5 gives the estimated partial correlation matrices when using 3 × 3 and 5 × 5 patches. The patches are extracted from all of the active regions and divided using the STARA and SMART masks into sunspots and magnetic fragments as before. The partial correlation of 3 × 3 patches is quite strong within both modalities. Based on a false alarm rate of 0.05, the theoretical thresholds for significance for the partial correlation (Hero & Rajaratnam 2011) of the 3 × 3 patches are approximately 0.0070 and 0.0014 for within the sunspots and magnetic fragments, respectively. For the 5 × 5 patches, the thresholds are 0.0080 and 0.0016, respectively. Given these thresholds, the partial correlation is statistically significant for nearly all values within the modalities (P_{xx} and ) using the 3 × 3 patches.
Fig. 5. Estimated partial correlation matrices of patch data from within the sunspots and the magnetic fragments using 3 × 3 (left) and 5 × 5 (right) patches. The theoretical thresholds (Hero & Rajaratnam 2011) for significance to attain a 0.05 false alarm rate are 0.0070 and 0.0014 for within the sunspots and magnetic fragments, respectively, when using a 3 × 3 patch. For the 5 × 5 patch, the thresholds are 0.0080 and 0.0016, respectively. Statistically insignificant values are set to zero. 
The crosspartial correlation when using the signed magnetic field (i.e. P_{xy} and P_{yx}) is very near zero in both regions (not shown). However, when we take the absolute value of the magnetogram patches, then the magnitude of the crosspartial correlation ( and ) is much higher in both regions suggesting that the correlation between the modalities is significant. The partial correlation is also stronger in magnitude in all cases within the sunspots than within the magnetic fragments.
The partial correlation matrices are very structured. In both sunspots and magnetic fragments, the pentadiagonallike structure within the modalities suggests that the image is generally stationary with approximately a thirdorder nearestneighbor Markov structure in the pixels. Such structure is clearly seen in the matrices for 5 × 5 patches. The crosspartial correlation also has a pentadiagonallike structure although the correlation is not as strong as within the modalities.
To better see the spatial correlations, in Figure 6 we plot the partial correlation patches taken from the columns of the sunspot partial correlation matrix corresponding to the center pixels when using 5 × 5 patches. Figure 6 shows clearly the greater partial correlation within the continuum. It also highlights that correlation is slightly higher in magnitude in the vertical direction than in the horizontal direction. Nearly all sunspots in this study are located within (−30°, +30°) from both the central meridian and the equator, and so projection effect is unlikely to cause this difference. The difference in correlation may be a feature of the sunspots themselves, but this may be difficult to determine since the difference in partial correlation is small.
Fig. 6. Partial correlation patches extracted from the columns in the sunspot partial correlation matrix corresponding to the center pixels. The partial correlation is stronger within the continuum. 
Some slight differences exist in the partial correlation matrices restricted to certain Mount Wilson classes. As an example, Figure 7 contains the partial correlation matrices within the sunspots after restricting the data to α and β groups as well as the difference between the absolute value of the two matrices. The α partial correlation matrix is higher in magnitude within the modalities than the β matrix but lower between the modalities. Within modalities, the strongest differences (a maximum of 0.056 and 0.067 within the continuum and magnetogram, respectively) are in the entries that correspond to pixels that are farther away from each other. In contrast, within the crosspartial correlation, the strongest differences (a maximum of 0.072) between the two AR types are in the entries that correspond to pixels that are close to each other. A similar pattern holds when comparing the α matrix to the matrices of the more complex groups.
Fig. 7. Partial correlation matrices within the sunspots using the data from α (a) and β ARs (b). Statistically insignificant values are again set to zero. (c) Difference between the absolute value of the α and β matrices. The α sunspots are more (resp. less) strongly correlated within (resp. between) the modalities than the β groups. 
Overall, the partial correlation matrices indicate that no larger than a 5 × 5 patch is necessary to capture the local spatial dependencies. A 5 × 5 patch of pixels corresponds roughly to the size of a mesogranule (Rieutord et al. 2000; Rast 2003). This suggests that within the magnetic fragments, it is likely that the granules and mesogranules within the photosphere contribute to the local spatial dependencies. Within the sunspots, a 5 × 5 patch corresponds to the size of the characteristic length of the largest penumbral filaments (Tiwari et al. 2013) which suggests that on average the local spatial dependencies are minimal beyond this scale. This analysis, however, does not rule out longrange spatial dependencies, which are more difficult to assess due to the large dimensionality. Future work will focus on this.
In the remainder of our analysis, we choose a 3 × 3 patch for the reasons mentioned in Section 2: to ensure that we capture the features of small sunspots and to limit the effects of high dimension on the accuracy of the analysis. Given these concerns, we see that 3 × 3 patches capture most of the spatial correlation. This is evident from Figure 5 where the partial correlation between pixels on opposite corners of a 3 × 3 patch is near zero and other pixels that are similarly far away from each other have low partial correlation. Thus a 3 × 3 patch strikes a good balance between scale, extrinsic dimension, and capturing the spatial correlation.
4.3. Canonical correlation analysis: methodology
To further investigate the correlation between the modalities, we use canonical correlation analysis (CCA) on the continuum patch x and the magnitude of the magnetogram patch y. CCA finds patterns and correlations between two multivariate data sets (Muller 1982; Nimon et al. 2010) and was used previously in the context of space weather e.g. for the combined analysis of solar wind and geomagnetic index data sets (Borovsky 2014).
In our application, CCA provides linear combinations of continuum patches x that are most correlated with linear combinations of magnetogram patches y. In other words, all correlations between the continuum and magnetogram patches are channeled through the canonical variables. Formally, CCA finds vectors a_{i} and b_{i} for i = 1, …, m^{2} such that the correlation is maximized and the pair of random variables and is uncorrelated with all other pairs u_{j} and v_{j}, j ≠ i. The variables u_{i} and v_{i} are called the ith pair of canonical variables while the vectors a_{i} and b_{i} are the canonical vectors. The solution a_{i} is the ith eigenvector of the matrix which is taken from the covariance matrix. The vector b_{i} is found similarly (Härdle & Simar 2007).
4.4. Canonical correlation analysis: results
Here we focus on 3 × 3 patches and apply CCA to all 424 image pairs using the magnitude of the magnetogram patches. Figures 8a and 8b show histograms of the estimated values of ρ_{1}. Within the sunspots, there are many groups with a near perfect correlation between the modalities and none of the groups have an estimated value below 0.41. Figure 8c plots the estimated values of ρ_{1} versus the number of samples used within the sunspots. Based on this plot, there are many ARs with high correlation and few patch samples suggesting that the correlation may be spurious. However, all of the estimated values are statistically significant as defined by the threshold given by Hero & Rajaratnam (2011) using a false alarm rate of 0.05 (shown as the magenta line in Fig. 8).
Fig. 8. Histograms of estimated ρ _{1} using CCA for within the sunspots (a) and the magnetic fragments (b) using 3 × 3 patches. (c) Scatter plot of ρ _{1} values and the number of samples available for within the sunspots. All points are above the magenta line which gives the threshold for statistical significance at a false alarm rate of 0.05 (Hero & Rajaratnam 2011). 
The histogram of ρ_{1} within the magnetic fragments (Fig. 8b) is quite different from the sunspot histogram (Fig. 8a). ρ_{1} is generally lower within the magnetic fragments than within the sunspots, which is consistent with the results in Figure 5. All of the ρ_{1} values are statistically significant.
The distributions of ρ_{1} differ slightly when comparing simple sunspot groups (α and β) with complex groups (βγ and βγδ). Figure 9 shows that complex groups generally have lower correlation between the modalities within the sunspots than the simpler groups. The estimated Hellinger distance (see Appendix A.3) between the distributions using the divergence estimator in Moon & Hero III (2014a, 2014b) is 0.22. Based on the central limit theorem of the estimator Moon & Hero III (2014b), this value is statistically significant with a pvalue of 1.6 × 10^{−12}. At least some of this difference is likely due to the smaller size of the simpler groups (and thus smaller sample size). However, it is unlikely to fully explain the difference given that there are many simple sunspot groups with high correlation and sufficient sample size.
Fig. 9. ρ _{1} histograms of complex (βγ and βγδ) and simple (α and β) regions within the sunspots (a) and the magnetic fragments (b). The simple ARs are generally more correlated within the sunspots but less correlated within the magnetic fragments. The difference between the sunspot distributions, as measured by the Hellinger distance, is statistically significant. 
Within the magnetic fragments, there are many more simple regions than complex regions with ρ_{1} < 0.4 (see the histogram in Fig. 9c). This could be related to the same phenomena that cause the intrinsic dimension to be higher within the magnetic fragments of simple sunspot groups observed in Section 3.3. However, the estimated Hellinger distance between these distributions is 0.016. Using the same statistical test, this estimate is not statistically significant with a pvalue of 0.31. Thus the distributions are not statistically different from each other.
To analyze the spatial patterns that produce the highest correlation between modalities, we apply CCA to the entire data set. Figure 10 plots ρ_{i} for i = 1, …, 9 for within the sunspots and within the magnetic fragments. The ρ_{i}s are all statistically significant. Notice that the ρ_{i}s are higher within the sunspots than the magnetic fragments which is consistent with the results in Figures 5 and 8.
Fig. 10. (a) Plot of the estimated ρ_{i} using CCA on the entire data set for i = 1, …, 9. All values are statistically significant (Hero & Rajaratnam 2011). (b) Canonical patches a_{i} (top) and b_{i} (bottom) for i = 1, …, 6 when using the entire data set from within the sunspots. The b_{i}s are approximately equal to the negative of the a_{i}s. 
Figure 10 shows the canonical patches a_{i} and b_{i} for i = 1, …, 6 when using all the data from within the sunspots. These are the spatial patterns within the two modalities that are most correlated with each other. The canonical patches have a “saddlelike” appearance where the gradient is positive in some directions and negative in others. For example, in a_{4}, the pixels to the left and right of the center are very negative but the pixels in the corners are all very positive. Note that these vectors correspond to centered values with respect to the mean patches.
Comparing the a_{i}s to the b_{i}s shows that the b_{i}s are approximately equal to the negative of the a_{i}s. This makes sense as sunspots within the continuum images correspond to a decrease in value relative to the background while ARs within the magnitude of the magnetogram images correspond to an increase in value relative to the background.
We also performed CCA separately on the data from the Mount Wilson classes. Figure 11 plots the ρ_{i} values for each class and the first canonical patches a_{1} and b_{1}. For ρ_{1} and ρ_{2}, the values for each class decrease in order of complexity (α, β, βγ, βγδ). This is consistent with our comparison of the partial correlation matrices in Figure 7 where the partial correlation was generally higher (in magnitude) for the α groups than the others. This is also consistent with the intrinsic dimension analysis in Section 3 where the intrinsic dimension generally increases with complexity. This is because if the correlation between and within modalities is higher, then fewer parameters are required to accurately describe the data which results in a lower intrinsic dimension.
Fig. 11. (a) Plot of the estimated ρ_{i} using CCA on data segregated by Mount Wilson classes for i = 1, …, 9 within the sunspots α groups start out with the highest correlation. (b) Canonical patches a_{1} (top) and b_{1} (bottom) for the Mount Wilson classes within the sunspots. Again, the b_{1}s are approximately equal to the negative of the a_{1}s as in Figure 10 but the patches differ slightly from class to class. 
The canonical patches a_{1} and b_{1} have similar patterns across the different classes although the patches for the βγ class are flipped compared to the others. The magnitude of the values in the βγδ patches is also smaller than that of the other patches.
Overall, the results of this section suggest that the two modalities are correlated in both the sunspots and the magnetic fragments and are therefore not independent. The correlation is stronger within the sunspots compared to the magnetic fragments and stronger within the sunspots in simple ARs compared to complex ARs. However, the correlation is not perfect and so there may be an advantage to including both modalities in the classification of sunspots and flare prediction.
5. Conclusion
Existing AR categorical classification systems such as the Mount Wilson and McIntosh schemes describe geometrical arrangements of the magnetic field at the largest length scale. In this work, we have focused on the properties of the ARs at fine length scale. We showed that when we analyze the global statistics or attributes of these local properties, we find differences between the simple and complex ARs as defined using the large scale characteristics. So by this approach, we are analyzing both the large and finescale properties of the images. Such results might be due to the multiscale properties of the magnetic fields, as evidenced previously in Ireland et al. (2008).
The local intrinsic dimension based on the kNN approach combines both continuum and magnetogram observations and provides some measure of local regularity for those images. Further differences between the Mount Wilson classes may be found by comparing the histograms or distributions of local intrinsic dimension of each individual AR instead of only comparing the means or pooled estimates as we did in this paper. There are several options to perform such comparisons. Each histogram could be treated as a vector, or we could consider the underlying probability density function within the framework of functional analysis. Supervised (using Mount Wilson classes) or unsupervised classification could be performed. Another option would be to view the set of histograms belonging to a specific class as samples from a distribution of vectors (or a distribution of probability density functions). Different classes could then be compared using divergence measures such as the Hellinger distance described in Appendix A.3.
This work also highlighted specific behaviors of the core of active regions (that corresponds to the sunspot masks in continuum) and magnetic fragments (the surrounding part of AR), as well as the difference of these two regions as a function of the Mount Wilson classification. We found that within the sunspots, the spatial and modal correlations are stronger than within the magnetic fragments. Additionally, simpler ARs were found to have higher correlation between the modalities within the sunspots than the complex ARs.
This study paves the way for further analysis based on dictionary learning. Knowledge of the intrinsic dimension allows us to choose the dictionary size. Moreover, the results of Section 3 showed that linear dictionary learning methods are sufficient. The spatial and modal correlation analysis in Section 4 justifies a choice of a patch size of 3 × 3 and confirms that both modalities (continuum and magnetogram) should be used in dictionary learning.
Acknowledgments
This work was partially supported by the US National Science Foundation (NSF) under Grant No. CCF1217880 and a NSF Graduate Research Fellowship to KM under Grant No. F031543. VD acknowledges support from the Belgian Federal Science Policy Office through the ESAPRODEX program, Grant No. 4000103240, while RDV acknowledges support from the BRAIN.be program of the Belgian Federal Science Policy Office, Contract No. BR/121/PI/PREDISOL. The editor thanks two anonymous referees for their assistance in evaluating this paper.
References
 Abramenko, V.I. Multifractal analysis of solar magnetograms. Sol. Phys., 228, 29–42, 2005, DOI: 10.1007/s1120700535259. [NASA ADS] [CrossRef] [Google Scholar]
 Bloomfield, D.S., P.A. Higgins, R.T.J. McAteer, and P.T. Gallagher. Toward reliable benchmarking of solar flare forecasting methods. Astrophys. J. Lett., 747, L41, 2012, DOI: 10.1088/20418205/747/2/L41. [NASA ADS] [CrossRef] [Google Scholar]
 Bornmann, P.L., and D. Shaw. Flare rates and the McIntosh activeregion classifications. Sol. Phys., 150, 127–146, 1994, DOI: 10.1007/BF00712882. [NASA ADS] [CrossRef] [Google Scholar]
 Borovsky, J.E. Canonical correlation analysis of the combined solar wind and geomagnetic index data sets. J. Geophys. Res. [Space Phys.], 119, 5364–5381, 2014, DOI: 10.1002/2013JA019607. [CrossRef] [Google Scholar]
 Bühlmann, P., and S. Van De Geer. Statistics for highdimensional data: methods, theory and applications. Springer Science & Business Media, Berlin, Germany, 2011. [CrossRef] [Google Scholar]
 Cadavid, A.C., J.K. Lawrence, and A. Ruzmaikin. Principal components and independent component analysis of solar and space data. Sol. Phys., 248, 247–261, 2008, DOI: 10.1007/s1120700790262. [CrossRef] [Google Scholar]
 Carter, K.M., R. Raich, and A.O. Hero. On local intrinsic dimension estimation and its applications. IEEE Trans. Sign. Process., 58 (2), 650–663, 2010. [CrossRef] [Google Scholar]
 Colak, T., and R. Qahwaji. Automated Mcintoshbased classification of sunspot groups using MDI images. Sol. Phys., 248, 277–296, 2008, DOI: 10.1007/s1120700790943. [NASA ADS] [CrossRef] [Google Scholar]
 Colak, T., and R. Qahwaji. Automated solar activity prediction: a hybrid computer platform using machine learning and solar imaging for automated prediction of solar flares. Space Weather, 7, S06001, 2009, DOI: 10.1029/2008SW000401. [NASA ADS] [CrossRef] [Google Scholar]
 Conlon, P.A., P.T. Gallagher, R.T.J. McAteer, J. Ireland, C.A. Young, P. Kestener, R.J. Hewett, and K. Maguire. Multifractal properties of evolving active regions. Sol. Phys., 248, 297–309, 2008, DOI: 10.1007/s1120700790747. [NASA ADS] [CrossRef] [Google Scholar]
 Conlon, P.A., R.T.J. McAteer, P.T. Gallagher, and L. Fennell. Quantifying the evolving magnetic structure of active regions. Astrophys. J., 722, 577–585, 2010, DOI: 10.1088/0004637X/722/1/577. [NASA ADS] [CrossRef] [Google Scholar]
 Costa, J.A., and A.O. Hero III. Determining intrinsic dimension and entropy of highdimensional shape spaces. Statistics and Analysis of Shapes, Springer, 231–252, 2006. [CrossRef] [Google Scholar]
 Dobigeon, N., J.Y. Tourneret, C. Richard, J. Bermudez, S. Mclaughlin, and A.O. Hero. Nonlinear unmixing of hyperspectral images: Models and algorithms. IEEE Signal Process Mag., 31 (1), 82–94, 2014. [CrossRef] [Google Scholar]
 Dudok de Wit, T., S. Moussaoui, C. Guennou, F. Auchère, G. Cessateur, M. Kretzschmar, L.A. Vieira, and F.F. Goryaev. Coronal temperature maps from solar EUV images: a blind source separation approach. Sol. Phys., 283, 31–47, 2013, DOI: 10.1007/s1120701201422. [CrossRef] [Google Scholar]
 Dudok de Wit, T., and F. Auchère. Multispectral analysis of solar EUV images: linking temperature to morphology. A&A, 466, 347–355, 2007, DOI: 10.1051/00046361:20066764. [CrossRef] [EDP Sciences] [Google Scholar]
 Efron, B., and R. Tibshirani. An Introduction to the Bootstrap. Chapman and Hall, Boca Raton, Florida, 1994. [Google Scholar]
 Elad, M., and M. Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Trans. Image Process., 15 (12), 3736–3745, 2006, DOI: 10.1109/TIP.2006.881969. [CrossRef] [MathSciNet] [PubMed] [Google Scholar]
 Gallagher, P.T., Y.J. Moon, and H. Wang. Activeregion monitoring and flare forecasting I. Data processing and first results. Sol. Phys., 209, 171–183, 2002, DOI: 10.1023/A:1020950221179. [NASA ADS] [CrossRef] [Google Scholar]
 Georgoulis, M.K. Turbulence in the solar atmosphere: Manifestations and diagnostics via solar image processing. Sol. Phys., 228, 5–27, 2005, DOI: 10.1007/s1120700525134. [NASA ADS] [CrossRef] [Google Scholar]
 Hale, G.E., F. Ellerman, S.B. Nicholson, and A.H. Joy. The magnetic polarity of sunspots. Astrophys. J., 49, 153–153, 1919, DOI: 10.1086/142452. [NASA ADS] [CrossRef] [Google Scholar]
 Härdle, W., and L. Simar. Applied multivariate statistical analysis. Springer, Berlin, Germany, 2007. [Google Scholar]
 Hero, A., and B. Rajaratnam. Largescale correlation screening. J. Am. Stat. Assoc., 106 (496), 1540–1552, 2011. [CrossRef] [Google Scholar]
 Hewett, R.J., P.T. Gallagher, R.T.J. McAteer, C.A. Young, J. Ireland, P.A. Conlon, and K. Maguire. Multiscale analysis of active region evolution. Sol. Phys., 248, 311–322, 2008, DOI: 10.1007/s1120700790280. [NASA ADS] [CrossRef] [Google Scholar]
 Higgins, P.A., P.T. Gallagher, R. McAteer, and D.S. Bloomfield. Solar magnetic feature detection and tracking for space weather monitoring. Adv. Space Res., 47 (12), 2105–2117, 2011. [NASA ADS] [CrossRef] [Google Scholar]
 Holappa, L., K. Mursula, T. Asikainen, and I.G. Richardson. Annual fractions of highspeed streams from principal component analysis of local geomagnetic activity. J. Geophys. Res. [Space Phys.], 119, 4544–4555, 2014, DOI: 10.1002/2014JA019958. [CrossRef] [Google Scholar]
 Ireland, J., C.A. Young, R.T.J. McAteer, C. Whelan, R.J. Hewett, and P.T. Gallagher. Multiresolution analysis of active region magnetic structure and its correlation with the mount wilson classification and flaring activity. Sol. Phys., 252, 121–137, 2008, DOI: 10.1007/s1120700892335. [NASA ADS] [CrossRef] [Google Scholar]
 Jolliffe, I.T. Principal component analysis. 2nd edn. SpringerVerlag New York, Inc., New York, 2002. [Google Scholar]
 Kestener, P., P.A. Conlon, A. Khalil, L. Fennell, R.T.J. McAteer, P.T. Gallagher, and A. Arneodo. Characterizing complexity in solar magnetogram data using a waveletbased segmentation method. Astrophys. J., 717, 995–1005, 2010, DOI: 10.1088/0004637X/717/2/995. [NASA ADS] [CrossRef] [Google Scholar]
 Kullback, S., and R.A. Leibler. On information and sufficiency. Ann. Math. Stat., 79–86, 1951. [CrossRef] [MathSciNet] [Google Scholar]
 Lauritzen, S. Graphical models. Clarendon Press, Oxford, ISBN: 9780191591228, 1996. [Google Scholar]
 Lawrence, J.K., A. Cadavid, and A. Ruzmaikin. Principal component analysis of the solar magnetic field I: the axisymmetric field at the photosphere. Sol. Phys., 225, 1–19, 2004, DOI: 10.1007/s1120700432572. [NASA ADS] [CrossRef] [Google Scholar]
 Levina, E., and P.J. Bickel. Maximum likelihood estimation of intrinsic dimension. Advances in Neural Information Processing Systems, 17, 777–784, 2004. [Google Scholar]
 Mallat, S., and Z. Zhang. Matching pursuits with timefrequency dictionaries. IEEE Trans. Signal Process., 41 (12), 3397–3415, 1993, DOI: 10.1109/78.258082. [NASA ADS] [CrossRef] [Google Scholar]
 Mayfield, E.B., and J.K. Lawrence. The correlation of solar flare production with magnetic energy in active regions. Sol. Phys., 96, 293–305, 1985, DOI: 10.1007/BF00149685. [CrossRef] [Google Scholar]
 McAteer, R.T.J., P.T. Gallagher, and P.A. Conlon. Turbulence complexity, and solar flares. Adv. Space Res., 45, 1067–1074, 2010, DOI: 10.1016/j.asr.2009.08.026. [NASA ADS] [CrossRef] [Google Scholar]
 McAteer, R.T.J., P.T. Gallagher, and J. Ireland. Statistics of active region complexity: a largescale fractal dimension survey. Astrophys. J., 631, 628–635, 2005, DOI: 10.1086/432412. [NASA ADS] [CrossRef] [Google Scholar]
 McIntosh, P.S. The classification of sunspot groups. Sol. Phys., 125, 251–267, 1990, DOI: 10.1007/BF00158405. [NASA ADS] [CrossRef] [Google Scholar]
 Moon, K.R., V. Delouille, J.J. Li, R. De Visscher, F. Watson, and A.O. HeroIII. Image patch analysis of sunspots and active regions. II. Clustering via matrix factorization. J. Space Weather Space Clim., 2015. [Google Scholar]
 Moon, K.R., and A.O. HeroIII. Ensemble estimation of multivariate fdivergence. Ensemble estimation of multivariate fdivergence. In Information Theory (ISIT), 2014 IEEE International Symposium on, IEEE, Honolulu, USA, pp. 356–360, 2014a. [Google Scholar]
 Moon, K.R., and A.O. HeroIII. Multivariate fdivergence estimation with confidence. Advances in Neural Information Processing Systems, 27, 2420–2428, 2014b. [Google Scholar]
 Moon, K.R., J.J. Li, V. Delouille, F. Watson, and A.O. HeroIII. Image patch analysis and clustering of sunspots: A dimensionality reduction approach. IEEE International Conference on Image Processing (ICIP), IEEE, Paris, France, 1623–1627, 2014. [Google Scholar]
 Muller, K.E. Understanding canonical correlation through the general linear model and principal components. Am. Stat., 36, 342–354, 1982. [Google Scholar]
 Nimon, K., R. Henson, and M. Gates. Revisiting interpretation of canonical correlation analysis: A tutorial and demonstration of canonical commonality analysis. Multivar. Behav. Res., 45, 702–724, 2010. [CrossRef] [Google Scholar]
 Phillips, K.J.H. Solar flares – a review. Vistas in Astron., 34, 353–365, 1991, DOI: 10.1016/00836656(91)90014J [CrossRef] [Google Scholar]
 Rast, M.P. The scales of granulation, mesogranulation, and supergranulation. Astrophys. J., 597 (2), 1200, 2003. [NASA ADS] [CrossRef] [Google Scholar]
 Rieutord, M., T. Roudier, J. Malherbe, and F. Rincon. On mesogranulation, network formation and supergranulation. A&A, 357, 1063–1072, 2000. [Google Scholar]
 Sammis, I., F. Tang, and H. Zirin. The dependence of large flare occurrence on the magnetic structure of sunspots. Astrophys. J., 540, 583–587, 2000, DOI: 10.1086/309303. [NASA ADS] [CrossRef] [Google Scholar]
 Scherrer, P.H., R.S. Bogart, R.I. Bush, J.T. Hoeksema, A.G. Kosovichev, et al. The solar oscillations investigation  Michelson doppler imager. Sol. Phys., 162, 129–188, 1995, DOI: 10.1007/BF00733429. [NASA ADS] [CrossRef] [MathSciNet] [Google Scholar]
 Stenning, D.C., T.C.M. Lee, D.A. van Dyk, V. Kashyap, J. Sandell, and C.A. Young. Morphological feature extraction for statistical learning with applications to solar image data. Statistical Analysis and Data Mining, 6 (4), 329–345, 2013, DOI: 10.1002/sam.11200 [CrossRef] [Google Scholar]
 Tiwari, S.K., M. van Noort, A. Lagg, and S.K. Solanki. Structure of sunspot penumbral filaments: a remarkable uniformity of properties. A&A, 557, A25, 2013. [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 Watson, F.T., L. Fletcher, and S. Marshall. Evolution of sunspot properties during solar cycle 23. A&A, 533, A14, 2011, DOI: 10.1051/00046361/201116655. [NASA ADS] [CrossRef] [EDP Sciences] [Google Scholar]
 Zharkova, V.V., S.J. Shepherd, and S.I. Zharkov. Principal component analysis of background and sunspot magnetic field variations during solar cycles 21–23. Monthly Notices of the RAS, 424, 2943–2953, 2012, DOI: 10.1111/j.13652966.2012.21436.x [CrossRef] [Google Scholar]
Appendix: Method details
A.1. Intrinsic dimension estimation of manifolds
Consider data that are described in an extrinsic Euclidean space of d dimensions. However, suppose the data actually lie on a lower dimensional manifold . Thus the intrinsic dimension m of the data corresponds to the dimension of . For example, data may be given to us in a 3dimensional space but lie on the surface of a sphere. Thus the intrinsic dimension of the data would be 2.
In some cases, data points from the same data set may lie on different manifolds. For example, part of the data with an extrinsic dimension of 3 could lie on the surface of a sphere (m = 2) while another part may lie on a circle (m = 1). We then say that data points from these different manifolds have a different local intrinsic dimension. The local intrinsic dimension gives some measure of the local complexity of the image. Additionally, the local intrinsic dimension is useful for dictionary learning because we can use it to determine whether differentsized dictionaries should be used for different regions, e.g. within the sunspots and outside of the sunspots.
We now describe the kNN estimator of intrinsic dimension in more detail. For a set of independently identically distributed random vectors Z_{n} = {z_{1}, …, z_{n}} with values in a compact subset of , the knearest neighbors of z_{i} in Z_{n} are the k points in Z_{n}\{z_{i}} closest to z_{i} as measured by the Euclidean distance  · . The kNN graph is then formed by assigning edges between a point in Z_{n} and its knearest neighbors. The intrinsic dimension is related to the total edge length of the kNN graph and can be estimated based on this relationship. The kNN graph is then formed by assigning edges between a point in Z_{n} and its knearest neighbors and has total edge length defined aswhere γ > 0 is a power weighting constant and is the set of k nearest neighbors of z_{i}. It has been shown that for large n,where α = (m − γ)/m, c is a constant with respect to α(m), and ε _{n} is an error term that decreases to 0 as n → ∞ (Costa & Hero III 2006). A global intrinsic dimension estimate is found based on this relationship using nonlinear least squares over different values of n (Carter et al. 2010).
A local estimate of intrinsic dimension at a point z_{i} can be found by running the algorithm over a smaller neighborhood about z_{i}. The variance of this local estimate is then reduced by smoothing via majority voting in a neighborhood of z_{i} (Carter et al. 2010).
A.2. Partial correlation
Let z be a random vector with size m. Let Σ be the covariance matrix of z, that is Σ_{ij} = Cov(z_{i},z_{j}), and let K = Σ^{−1} be the inverse of the covariance matrix, also called the precision matrix.
The partial correlation between z_{i} and z_{j} given all the other variables z\{z_{i},z_{j}} measures the degree of correlation between these two variables after removing the effect of the remaining ones. Let P_{ij} denote the partial correlation between z_{i} and z_{j}. It has been shown (Lauritzen 1996) that P_{ij} can be related to the elements of the precision matrix K as follows:
A.3. Estimating the Hellinger distance
Information divergences are a class of functionals that measure the difference between two probability distributions. The most popular divergence measure is the KullbackLeibler divergence (Kullback & Leibler 1951). The Hellinger distance is another divergence measure and is defined aswhere f and g are the two probability densities being compared. The Hellinger distance is a metric which is not true of divergences in general. we use the nonparametric divergence estimator derived in Moon & Hero III (2014a, 2014b). In Moon & Hero III (2014a, 2014b), it was shown that this estimator converges to the true divergence with mean squared error convergence rate O(1/T) where T is the number of samples from each probability distribution. In Moon & Hero III (2014b), it was shown that the distribution of the normalized version of this estimator converges to the standard normal distribution. We can use this fact combined with a bootstrap estimate (Efron & Tibshirani 1994) of the variance of the estimator to test the hypothesis that the divergence is zero (and hence the distributions are equal).
Cite this article as: Moon KR, Li JJ, Delouille V, De Visscher R, Watson F, et al. Image patch analysis of sunspots and active regions. I. Intrinsic dimension and correlation analysis. J. Space Weather Space Clim., 6, A2, 2016, DOI: 10.1051/swsc/2015044.
All Tables
Number of each AR per Mount Wilson class. Simple ARs include α and β groups while complex ARs are βγ and βγδ groups.
Estimated intrinsic dimension results for different groups of ARs in the form of mean ± standard deviation. The complex ARs have higher intrinsic dimension within the sunspots than the simple ARs but lower intrinsic dimension within the magnetic fragments.
All Figures
Fig. 1. An example patch from the edge of a sunspot in a continuum image and its column representation. 

In the text 
Fig. 2. Examples of the estimated local intrinsic dimension using the kNN method for an α group (top) and a βγδ group (bottom). Regions with more spatial structure have lower intrinsic dimension. 

In the text 
Fig. 3. Normalized histograms of mean estimated intrinsic dimension of α, β, βγ, and βγδ groups using the kNN method. The distributions of intrinsic dimension differ by complexity with simpler AR groups having higher (resp. lower) intrinsic dimension within the sunspot (resp. magnetic fragments). 

In the text 
Fig. 4. Normalized histograms of pooled local estimates of intrinsic dimension of α, β, βγ, and βγδ groups using the kNN method. The distributions of intrinsic dimension differ by complexity with simpler AR groups having higher (resp. lower) intrinsic dimension within the sunspot (resp. magnetic fragments). 

In the text 
Fig. 5. Estimated partial correlation matrices of patch data from within the sunspots and the magnetic fragments using 3 × 3 (left) and 5 × 5 (right) patches. The theoretical thresholds (Hero & Rajaratnam 2011) for significance to attain a 0.05 false alarm rate are 0.0070 and 0.0014 for within the sunspots and magnetic fragments, respectively, when using a 3 × 3 patch. For the 5 × 5 patch, the thresholds are 0.0080 and 0.0016, respectively. Statistically insignificant values are set to zero. 

In the text 
Fig. 6. Partial correlation patches extracted from the columns in the sunspot partial correlation matrix corresponding to the center pixels. The partial correlation is stronger within the continuum. 

In the text 
Fig. 7. Partial correlation matrices within the sunspots using the data from α (a) and β ARs (b). Statistically insignificant values are again set to zero. (c) Difference between the absolute value of the α and β matrices. The α sunspots are more (resp. less) strongly correlated within (resp. between) the modalities than the β groups. 

In the text 
Fig. 8. Histograms of estimated ρ _{1} using CCA for within the sunspots (a) and the magnetic fragments (b) using 3 × 3 patches. (c) Scatter plot of ρ _{1} values and the number of samples available for within the sunspots. All points are above the magenta line which gives the threshold for statistical significance at a false alarm rate of 0.05 (Hero & Rajaratnam 2011). 

In the text 
Fig. 9. ρ _{1} histograms of complex (βγ and βγδ) and simple (α and β) regions within the sunspots (a) and the magnetic fragments (b). The simple ARs are generally more correlated within the sunspots but less correlated within the magnetic fragments. The difference between the sunspot distributions, as measured by the Hellinger distance, is statistically significant. 

In the text 
Fig. 10. (a) Plot of the estimated ρ_{i} using CCA on the entire data set for i = 1, …, 9. All values are statistically significant (Hero & Rajaratnam 2011). (b) Canonical patches a_{i} (top) and b_{i} (bottom) for i = 1, …, 6 when using the entire data set from within the sunspots. The b_{i}s are approximately equal to the negative of the a_{i}s. 

In the text 
Fig. 11. (a) Plot of the estimated ρ_{i} using CCA on data segregated by Mount Wilson classes for i = 1, …, 9 within the sunspots α groups start out with the highest correlation. (b) Canonical patches a_{1} (top) and b_{1} (bottom) for the Mount Wilson classes within the sunspots. Again, the b_{1}s are approximately equal to the negative of the a_{1}s as in Figure 10 but the patches differ slightly from class to class. 

In the text 
Current usage metrics show cumulative count of Article Views (fulltext article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 4896 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.