full-disk Hα images. The initial step is to detect the filaments in the solar image, then determine if they are growing, stable or disappearing. If a disappearing ...
Development of an Automatic Filament Disappearance Detection System Jianlin Gao1 , Haimin Wang2 and Mengchu Zhou 1 1
Department of Electrical and Computer Engineering, New Jersey Institute of Technology, Newark, NJ 07102, USA 2 Department of Physics, New Jersey Institute of Technology, Newark, NJ 07102, USA
Abstract. This paper presents an efficient and automatic method for detecting filament disappearances. This method was applied to the Big Bear Solar Observatory's (BBSO) full-disk Hα images. The initial step is to detect the filaments in the solar image, then determine if they are growing, stable or disappearing. If a disappearing filament is found, the solar community can be automatically alerted in near real time. This system is proven to be accurate and fast. In addition, a study of the appearance and disappearance of all filaments in 1999 is presented.
1. Introduction
The study of solar filaments is a very important subject in solar physics and it is closely related to the study of space weather. Filaments are amazing objects. They are located in the corona but have temperatures only one one-hundredth of the corona and densities one hundred times greater than the local corona values. In Hα light, filaments are dark ribbons against the bright solar disk. At the limb, they appear bright against the dark sky and are called prominences. Martin (1998) gives a comprehensive review of observational conditions for the formation and stability of filaments. Dumitrache (1997) describes the evolution of filaments.
1
This paper describes a new solar filament detection algorithm applied to full-disk Hα images from the Big Bear Solar Observatory (BBSO). BBSO typically produces approximately 600 full-disk Hα images per day. Denker et al. (1999) describes the instrument that takes the full-disk Hα images, the data calibration procedures and image processing techniques that are applied to the daily full-disk Hα images. To use these images to detect disappearing filaments, it is desirable to use an automated detection system. This system has the following advantages: it is fast, accurate and the cost of documenting erupting filaments can be reduced by use of a computer program. Wang et al. (1998) analyzed disappearing filaments during the period from September 1991 through September 1994 and made two statistical studies. The first study is the latitude distribution of all large disappearing filaments. The second study is concerned with the disk distribution of all detectable disappearing filaments. For these two studies all events were detected manually. The primary purpose of the system is to detect filaments accurately and quickly. We developed the program in Interactive Data Language (IDL) by Research Systems, Inc. The program runs on a Pentium II personal computer operating Linux. A shell program automatically runs the filament detection program and emails the results to interested users. The specifics of the filament detection algorithm are discussed below.
2. Filament Disappearance Detection Algorithm
This paper applies a method of image segmentation. The goal is to detect regions on the sun that satisfy certain predefined homogeneity criteria. Starting at some seed location,
2
neighboring pixels are examined, one at a time, and added to the region if they are sufficiently similar. The procedure continues until all adjacent pixels have been tested [Besl, 1988; Beveridge, 1989]. The algorithm discussed in this paper combines thresholding and region based techniques. In the full-disk solar image, the candidate filament pixels can be classified using thresholding. The region growing method merges the candidate pixels into filament areas. Because filaments are dark in nature, we set a threshold value to select the filaments. We use the median of the intensity of the entire image as a reference level since it is less sensitive to asymmetric intensity distributions than the mean value. The threshold value is typically set to 50% of that median value. The solar images are recorded by a KODAK Megaplus Model 4.2 CCD camera which has an array of 2032 × 2028 pixels. Images are scaled to 1024 × 1024 by the filament detection program. The detection algorithm runs quickly despite the large size of the data array. Each image typically can be processed in less than one minute.
2.1
The filament detection algorithm
We developed an algorithm that runs fast and detects filaments accurately. A flow chart of the algorithm is shown in Figure 1.
3
Acquire data
Set the pixels under intensity threshold to 1, otherwise 0
For all pixels inside solar limb
Pixel value=1?
Set the processed pixels value to 2
N
Y N The adjacent pixels value =1?
Y
Out of solar limb?
Y
N N
Is the total size larger than the set value? Y Record the filament
N Finished? Y End
Figure 1. The flow chart of an efficient algorithm for filament detection
4
As mentioned above, four important issues should be considered in the algorithm: (1) Thresholds There are two different threshold values in the algorithm. The first threshold is used to detect the dark filament pixels by their intensity values. This threshold will be referred to as the "intensity threshold" throughout this paper. Another threshold is used to determine if a detected dark area is large enough to be regarded as a filament by the total number of dark pixels in the region. This threshold will be referred to as the "size threshold" in this paper. (2) Dark areas outside the limb In the process of detecting a filament, the adjacent pixels may be out of the solar limb. It is possible that a pixel outside the solar limb can be mistaken as part of a filament. In the program, one additional condition is set. If the filament area under consideration touches the solar limb the pixel will be deleted. This deletion does not affect our tracking disappearing filaments because we consider the filaments on the disk of the solar image more than those near the solar limb. When a filament appears beyond the solar limb, it appears as a bright structure known as a prominence. At this time, our program does not detect prominences. (3) Adjacent pixel connection A "4-pixel distance" method is used to connect the adjacent pixels in the detection of individual dark areas. The method is shown in Figure 2. In Figure 2 (a), only the eight pixels that are immediately adjacent to the candidate pixel under consideration are checked during region growing. Because of the presence of error pixels, additional
5
"adjacent" pixels should be checked. Thus, so called "bridge pixels" must be added to allow such filaments to be correctly detected. The "4-pixel distance" method is applied to the data to repair the error pixels in the solar images. Figure 2 (b) shows the "4-pixel distance" method. A total of eighty adjacent pixels to the candidate pixel are checked with this procedure.
(a) one pixel distance
(b) four pixels distance
The current filament pixel
The adjacent pixels
Figure 2. The adjacent pixels connection
(4) In some solar images, a single large filament may be broken into several smaller filaments by error pixels. For example, the actual area of a filament may be 300 pixels. If it is separated into three smaller areas that have 150, 100 and 50 pixels respectively, the filament will be considered to be noise since all three areas are under the "size threshold" of 220 pixels. To ensure the program detects this area as one large filament, these error pixels must be detected and repaired. If the total distance between two adjacent filaments is less than 40 pixels, the two separately detected filaments are considered the same, larger filament. This method is referred to as the "40-pixel distance" method.
6
If the "40-pixel distance" method is independently applied to the solar images, many error filaments will be detected because of the many error pixels. Therefore, the "4pixel distance" method is used to detect the individual small filaments and the "40-pixel distance" method is used to determine if two or more adjacent filaments are actually one large filament. Unfortunately, the "40-pixel distance" method runs very slowly and would take longer to process a single image than our requirements demand. Therefore, the program is most efficient when both the "4-pixel distance" and "40-pixel distance" methods are used together. The "intensity threshold" value is selected to be half of the median intensity value of the solar image. Any pixel whose intensity value is over this "intensity threshold" value is set to 1. Otherwise, the pixel value is set to zero. The algorithm begins with one reference pixel whose intensity is under the "intensity threshold". Each adjacent pixel is then checked. If at least one adjacent pixel is below the "intensity threshold", the pixel is added to the region and is regarded as the current reference pixel. The adjacent pixels of the new reference pixel are then checked. The process continues until the area's adjacent pixels are all above the "intensity threshold". The total number of pixels in the area is then obtained. If the total exceeds the "size threshold" value, it is considered a filament. The pixel intensity value of this area is then set to 255. If the total number of pixels in the area under consideration is below the "size threshold", the area is considered to be not a filament and its pixel intensity values are set to 2. As the program continues to check for additional filaments on the solar disk, these pixels are no longer tested. The actual intensity and "size threshold" levels for the program were determined by our tests on the Big Bear data.
7
2.2
Filament disappearance detection After the filaments are detected in individual images, the results from two
consecutive images are compared to detect filament disappearances. Taking into account the solar rotation of the sun during the time between the two images under consideration, the location of filament disappearances can be detected. Our first implementation was to select one image per day to detect filament disappearances during that particular two days. The following issues are important in the process of filament disappearance detection. (1) Solar rotation Solar rotation can carry a filament over the west limb of the sun and out of our view without any physical changes in the filament's structure. Therefore, our program must take this effect into account and not log a filament that rotates out of view as a disappearing filament. (2) Position of the sun on the detector Due to slight errors in telescope tracking, the position of the image of the sun on our detector can change by a small amount. If not taken into account, this effect can mimic a filament disappearance. It is very important that the solar community be alerted to filament disappearances in a timely manner. To accomplish this task, the program periodically downloads all recent full-disk Hα images, detects all filaments, categorizes them into new, partially
8
disappeared, fully disappeared, stable and growing and alerts the solar community, via email, should any relevant events be detected.
3. Result 3.1 Filaments detected Figure 3 presents one example of filament detection and filament disappearance detection. To ensure that the smallest filaments will be detected, the "size threshold" of the filament areas was set to 205 arcsec2 . While the "size threshold" of disappearing filament is set to 750 arcsec2 . The filament disappearances are reported in Table 1.
Event Number 1 2 3 4
Position S13 E32 S05 E11 N30 E08 S04 W02
Area ( Arcsec2 ) 1211 1318 1184 2009
Table 1. Filament disappearances between 17:02 Feb. 22, 1999 and 16:21 Feb. 23, 1999
All of the available solar images from 1999 have been analyzed for this paper. Three statistical studies of filaments and detected filaments disappearing, (1) size distribution (2) filament frequency in time series and (3) latitude distribution, based on this set of data are presented. Our intention was not to present new results but to check that our filament detection algorithm is working correctly. The three statistical results are discussed below.
9
(a) Feb. 22, 1999 Solar image and filaments detected
(b) Feb. 23, 1999 solar image and filaments detected
(c) Filament disappearances between Feb. 22 and Feb. 23 1999 Figure 3. One example of filament disappearance detection
10
3.2 Size distribution of filament disappearance areas In Figure 4, the size distribution of all the detected filaments and filament disappearances are shown. The filament "size threshold" was set at 750 Arcsec2 . Using a curve fitting technique, the relationship between size distribution (N) and the filament disappearance area (S) is as follows: N = K • Sα where K value is a function of the "size threshold" selection. N and S follow a power law with a power index of –2.6. Wang et al. (1998) analyzed the relationship between the size distribution and the length of filaments that disappeared. They found that the distribution follows a law with a power index of –1.4. However, they only used the one-dimensional length distribution. It is reasonable to conclude that a two-dimensional distribution has a steeper power spectrum shape, because longer filaments appear to be proportionally wider than shorter filaments. As a comparison, the flux distribution of flare emissions in hard X-rays follows a power-law distribution, with a power index of –1.8 (Crosby, Ashwanden, and Dennis, 1993).
11
Total disappearances in Year 1999
3
10
2
10
1
10
0
10 2 10
3
10
4
10
Size of disappeared filaments (Arcsec2 )
Figure 4. Size distribution of filaments disappeared
12
3.3 Filament frequency in time series
4
Number
3.5 3 2.5 2 1.5 1 1
2
3
4
5
6
7
8
9
10 11 12
Month of 1999 All detectedl filaments
(a)
Filaments disappeared
(b)
Figure 5. The number of filaments and filaments disappeared
The number of all detected filaments as a function of time in 1999 is shown in Figure 5 (a). While the number of disappearing filaments as a function of time in 1999 is shown in Figure 5 (b). The number of filaments peak in January/February and August/September. The time difference between these two peaks is about six months. In Figure 5 (b), the number of filaments that disappeared, as a function of the time, is shown. More filaments disappeared in March, August and September and the time between the two most active months is five to six months. Bogart and Bai (1985) presented a 152-day periodicity in the occurrence of solar flares. We may also have detected such a periodicity in the occurrence
of
filaments,
although
only
13
two
complete
periods
are
shown.
Number
3.4 Latitude distribution of filaments and filaments disappeared
100 90 80 70 60 50 40 30 20 10 0 -90
-60
-30
0
30
60
90
Latitude All detected filaments
Filaments disappeared
Figure 6. Distribution of filament positions
We also analyzed the latitude at which the filaments in our study occurred. In Figure 6, we plot the distribution of filament midpoint latitudes. We note that the location of filaments peaks at approximately ± 30 degrees. This corresponds very well with the sun's active latitudes.
4. Conclusion and future work
14
An efficient algorithm to automatically detect solar filaments is introduced in this paper. This algorithm plays a key role in the detection of disappearing filaments at BBSO. "Intensity threshold" and region growing methods are combined in the algorithm. "Intensity threshold" is used to detect the dark pixels that are filament candidates. The region growing method merges the detected pixels to form the filament areas. Automatic filament detection is an important first step to real time filament disappearance detection. BBSO records one solar image every minute for a total of typically 600 per day. Due to the high running speed of the algorithm, we can process at least one image every minute. A shell program has been developed to automatically processes the images. All full disk Hα images from 1999 have been analyzed to demonstrate the statistical results. The size distribution of filament disappearances follows a power law with a power index of –2.6. The time between the two most active periods of filament disappearances is about five to six months. The solar disk location of filaments peak at approximately ± 30 degrees. Obviously, if a filament disappearance occurs near the solar disk center, it has the potential to generate a major geomagnetic event. Our results are potentially useful in establishing a quantitative relationship between filament eruptions and geomagnetic activity. In addition, the latitude distribution and solar cycle variation may provide valuable information for the study of the solar dynamo. The future work is to process high resolution solar images from BBSO with this proposed algorithm. More accurate locations of filament events can be reported.
15
Acknowledgements This work is supported by NSF under grant ATM-9713359, ONR under grant N00014-9711037, and NASA under grant NAG5-7085.
References Beveridge, R.: 1989, Computer Vision, Graphics, Image Processing. 2, 311. Bogart, R. S. and Bai, T.: 1985, Astrophysical J. 299, L51 Castleman, K. R.: 1996, Digital Image Processing, Prentice Hall, Englewood Cliffs, New Jersey. Chen, S., Lin, W. and Chen, C.: 1991, Computer Vision, Graphics, Image Processing: Graph, Models Image processing. 53, 457. Crosby, N. B., Aschwaden, M. J., and Dennis, B.R.: 1993, Solar Physics. 143, 275. Denker, C., Johannesson, A., Marquette, W., Goode, P., Wang, H. and Zirin, H.: 1999, Solar Physics. 184, 87. Dumitrache, C.: 1997, Solar Physics. 173, 281. Martin, S. F.: 1998, Solar Physics. 182, 107. Russ, J. C.: 1992, The Image Processing Handbook. CRC Press. Wang, H., Komenda, A.E., Tand, F. and Zirin, H.: 1998, Solar Physics. 178, 109.
16
