Документ взят из кэша поисковой машины. Адрес оригинального документа : http://hea-www.harvard.edu/AstroStat/Stat310_1011/ds_20110329.pdf
Дата изменения: Tue Mar 29 16:40:57 2011
Дата индексирования: Tue Oct 2 04:18:35 2012
Кодировка:

Automatic Classification of Sunspot Groups Using SOHO/MDI Magnetogram and White Light Images
David Stenning
March 29, 2011

1

Introduction and Motivation

Although solar data is being generated at an unprecedented rate, the majority of sunspot classification is done manually by experts. In addition to being subject to observer bias, manual classification is labor-intensive and time-consuming. Therefore, a fast and automatic classification procedure is highly desirable.
2

Mount Wilson Classification

Using the Mount Wilson scheme, sunspot groups can be classified into the following four broad categories (from Ireland et al. 2008):

Alpha: a single dominant spot often linked with a plage of opposite magnetic polarity Beta: a pair of dominant spots of opposite polarity Beta-Gamma: bipolar groups with more than one clear North-South polarity inversion line Beta-Gamma-Delta: bipolar groups with umbrae of opposite polarity together in a single penumbra
3

4

Extracting Features From Magnetograms

As a first step, magnetograms are "cleaned" by a morphological opening operation (see Soille, 2003) with a spherical structuring element of radius 2. This is to smooth the "white" areas in the magnetograms.

Left: Original Magnetogram

Right: "Opened" magnetogram
5

This is repeated on the inverse magnetogram image to smooth the black regions.

Left: Inverse Magnetogram

Right: "Opened" Inverse Magnetogram
6

Using the cleaned images, sunspot active-region "seeds" are obtain by thresholding. In this operation, the threshold value is set to be 2.5 sample standard deviations above the sample mean (calculated from all pixel values in the image). All pixels above this value are marked as belonging to the active-region.

Left: Original Magnetogram

Right: Active-region seeds

7

Using these seeds, we can extract features:
Ratio: The ratio of active-region seeds. A(w) :

Let W denote the set of all white active-region seeds. Compute the center of mass, c, of W. For each w in W, walk from w to c in a straight line counting (i) the total number of pixels traveled and (ii) how many pixels are An W (i.e. are part of the white active i w region). Denote (i) by L(w) and (ii) by l(w). Then A(w) is given by

where |W| is the number of pixels in W. B(w): Computed the same as A(w) but with the black activeregion seeds.
8

Since we are interested in the separation of the two polarities, we use seeded region growing (see Adams and Bischof, 1994) to obtain the separating line.

Magnetogram

Active-Region Seeds

Binary Regions

The curvature of the separating line is computed by averaging second derivatives on the line designated by where the white and black regions meet. This is the the third feature, denoted by curve.

9

To detect delta spots, we need to be able to compare polarities of umbrae within a single penumbra in sunspots. To do this, we need to utilize the white-light (intensity-gram) images. Our first step, as with the magnetograms, is to clean the image with a morphological opening:

Left: White-Light Image

Right: Cleaned White-Light Image

10

Using th applied. that are only the umbra.

e cleaned image, a two-stage thresholding operation is First, the entire image is thresholded to obtain pixels part of the sunspot. Then, thresholding is applied on sunspot pixels to identify pixels that are part of an

Left: Original WL Image

Right: Penumbra and Umbra Regions

We can then examine each closed penumbra region to determine if there are umbrae of different polarities within a single penumbra. The outcome is denoted by the categorical variable delta (1 if true, 0 otherwise).

11

Features

A(W): "Scatter" of white active-region pixels B(W): "Scatter" of black active-region pixels Ratio: The ratio of white-to-black or black-to-white active-region pixels (depending on which active-region has more pixels associated with it). Curve: The curvature of the separating line Delta: Indicator as to whether there are umbrae with different polarities within one penumbra.

12

Using "Good" Examples

To test the features we have created, 128 sunspot groups were selected to be "good" examples of the 4 classes. They are broken down as follows:

28 66 22 12

Alpha Be t a Beta-Gamma Beta-Gamma-Delta

13

Exploratory Data Analysis

14

15

16

17

18

19

20

21

22

23

Delta Spots
Number of images identified as having delta spots (delta=1) by class:

A B B B

lpha: 0 out of 28 eta: 0 out of 66 eta-Gamma: 10 out of 22 eta-Gamma-Delta: 7 out of 12

24

A beta-gamma active-region that was identified as having a delta spot:

25

The same region 24 hours later, now identified as betagamma-delta:

26

A beta-gamma-delta active-region that was not identified as having a delta spot:

27

Off the Shelf Classifier: Multinomial Logistic Regression
Results from multinomial logistic regression using all available features:
Actual Alpha Predicted Alpha Beta B-G B-G-D 17 11 0 0 Beta 4 62 0 0 B-G 3 9 8 2 B-G-D 2 3 6 1

28

Off the Shelf Classifier: Multinomial Logistic Regression

Results from multinomial logistic regression using all available features, including interactions and quadratic terms:
Actual Alpha Predicted Alpha Beta B-G B-G-D 19 9 0 0 4 62 0 0 3 7 9 3 1 3 3 5 Beta B-G B-G-D

29

Next Steps

Tweak features to obtain better separation between the classes

Examine the cases where the delta classifier is failing. Since many beta-gammas are identified as having delta spots (and many beta-gamma-deltas were not identified as such), perhaps modify the delta feature to make it more continuous.

Consider quadratic discriminant analysis and other methods to obtain more complex decision boundaries.

30

Acknowledgements
This work is joint with:

V. Kashyap T. C. M. Lee D. van Dyk C. A. Young

31

References

R. Adams and L. Bischof, "Seeded region growing," IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, vol. 16, no. 6, pp. 641647, 1994. [Online]. Available: http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=295913 J. Ireland, 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 mt. wilson classif cation and f aring activity," Solar i l Physics, 2008. [Online]. Available: http://arxiv.org/abs/0805.01 P. Soille, Morphological Image Analysis, P. Soille, Ed. 2003, vol. 132.01 Springer-Verlag,

32