Sunspot Identification and Tracking with OpenCV

Ruben du Toit Günther Drevin

School of Computer Science and Information Systems School of Computer Science and Information Systems
North-West University North-West University
Potchefstroom, South Africa Potchefstroom, South Africa
[email protected] [email protected]

Nicolaas Maree Du Toit Strauss

School of Computer Science and Information Systems School of Physical and Chemical Sciences
North-West University North-West University
Potchefstroom, South Africa Potchefstroom, South Africa
[email protected] [email protected]

Abstract—Space weather forecasting has gained interest over In 1611 Goldsmid, Galilei, Scheiner and Harriot [2] were
the past five decades since mankind started exploring space and the first to observe these sunspots through a telescopes.
began using highly developed technology. Dangers such as coronal Goldsmid was the first to publish his finding and also came
mass ejections (CME) can negatively affect the technology on
which modern society relies heavily for communication and up with the theory of the Sun’s rotation due to the sunspots
navigation. This study uses continuum images of the Sun to movement on its disk. In 1630 it was noted that the sunspots
identify sunspots which lead to CMEs and track them over time. appear only within low altitudes zones on either side of the
The study consisted of two phases where the first was to identify equator.
every sunspot on a continuum image by using OpenCV methods
in Python. The second phase tracked those sunspots over a video A significant discovery was made by Schwabe in 1851 that
consisting of continuum images. The most important part of the recorded the periodicity of sunspots and their occurrence every
study was to correctly identify each sunspot on the image, for
11 years. A number of years later Carrington determined that
it determined the rest of the study’s results. The key method
used in the identification phase was the Canny edge detection the Sun rotates more rapidly at the equator and slower at higher
method that identified all the edges of the sunspots. The tracking altitudes.
main method made use of the multi-object tracker function
of OpenCV to track multiple sunspots over the video. In this Sunspots are dark areas that grow, and decay and these
study, background is provided to help understand the reasons regions are located on the lowest layer of the Sun that emits
for developing such a system as well as the results obtained. the light we see on Earth known as the photosphere [3].
The results obtained using the above methods were promising, These sunspots are visible due to their trait for being cooler
indicating promising future work into this field and project.
Index Terms—sunspot detection, tracking, image processing
than their surrounding area on the solar surface causing them
to be darker. Sunspots start out as small dark spots on the
I. I NTRODUCTION photosphere named pores. These pores either decay early in
their life cycle or develop into the sunspots that are referred to
In today’s world, society relies on communication and
in this study. As these sunspots start to enlarge, small portions
power systems to keep the world turning, but these systems are
tend to break away from the original sunspot. This occurrence
often interrupted by things that aren’t visible to the naked eye
leads to the classification of groups and the counting of
[1]. Space weather plays a major role in keeping these systems
sunspots. When a sunspot becomes large enough it forms a
fully functional by executing numerous tasks that involves the
double-ended group and if viewed with a solar magnetograph
Sun and its sunspots. There exist several manual systems that
it can be determined that opposite sides of the groups have
can detect, track and classify sunspots, but not all systems use
opposite magnetic polarity. Using the solar magnetograph
the same techniques to achieve their maximum efficiency. In
instrument, the sunspot groups can be defined as the collection
this section, the background on important terms and existing
of sunspots.
systems will be discussed to give insight into this study.
In the early 1600s, Chinese astronomers were the first to In this study, existing sunspots detection systems will be
report the sighting of dark spots on the Sun that later became briefly discussed and methods that are used to develop these
known as sunspots [2]. This observation was possible due to systems. To develop an automated sunspot detection system
the Sun having the optimal conditions in terms of being just without any manual input could help in identifying and clas-
above the horizon where it is possible to view the Sun with sifying these sunspots where human error could occur, which
the naked eye. could lead to potential damage of technology.
978-1-7281-4162-6/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: NED UNIV OF ENGINEERING AND TECHNOLOGY. Downloaded on August 07,2024 at 06:07:18 UTC from IEEE Xplore. Restrictions apply.
 II. C ORONAL M ASS E JECTIONS as these images have a time difference of approximately thirty
Coronal mass ejections are similar to solar flares in terms of minutes. Also, the size of the images differ. The magnetogram
having large eruptions of magnetic field and plasma into the images must, therefore, be realigned, correlated and resized,
heliosphere. According to Carley et al. [4], these eruptions to ensure that both the magnetogram and continuum images
carry energetic electrons which produce various sources of both have the same center, solar coordinates and radius. This
high intensity plasma emissions. When solar flares occur on process is done by mapping the magnetogram image from
sunspots, forms of light, such as X-rays and radio waves heliocentric-Cartesian coordinates to Carrington Heliographic
take about 8 minutes to reach the Earth from the Sun where coordinates. This step is repeated in reverse where the image is
they are not often dangerous to the Earth, rather they act again re-mapped to the Heliocentric-Cartesian coordinates due
as indicators of potential coronal mass ejections. After solar to the initiation of the merging of the two images. The merging
flares occur, the likelihood of coronal mass ejections occurring of the two images uses the solar coordinates, center and radius
is highly possible and this solar activity is what causes of the continuum image as the new values for the magnetogram
disruptions in communication and navigation systems, due to image. In Figure 1, A represents the magnetogram image and
magnetic storms that hit the Earth and carry charged particles. B the continuum image. C represents the result of Stage 1
According to [1] these storms can affect and damage not and Stage 2 processing on the magnetogram. D represents the
only satellites that provides wireless communication from continuum image result after Stage 1 processing.
Solar feature procedures can be divided into the following
space, but also ground systems, such as railway signalling,
seven categories ranging from the simplest to the most com-
high voltage power transmission grids and telecommunications
plex:
cables. These CMEs can take a few hours to a few days to
reach the Earth [1]. In the next section, some of the existing 1) Threshold approaches: The threshold approaches are
sunspot detection and tracking systems will be discussed. derived from image histograms that has its features
superimposed on a constant background intensity.
2) Edge based methods: Applies either first or second
derivatives of the spatio-temporal intensity function but
does tend to produce incomplete edges which then need
linking when applied to solar images.
3) Region growing methods: This method requires a start-
ing pixel for merging with adjacent pixels, if it meets
the correct criterion.
4) Hybrid region growing and edge-based techniques: This
method comes with the cost of longer processing times
and higher complexity. This is due to it using both edge
detection properties and homogeneity of regions into the
segmentation process for improved results.
5) Artificial neural network methods: A modern technique
used to train and recognise features that are not in a
given criteria. Once it is trained it is fast and accurate
Fig. 1. Pre-processing technique example in its execution of its given task.
6) Global information methods: This method returns very
good results and feedback, but just as category 4, is very
time-consuming in execution and very complex.
III. I MAGE P ROCESSING TECHNIQUES
7) Miscellaneous methods: Includes all other methods not
Pre-processing is the process of detecting and extracting the mentioned in the previous categories, including data
sunspots in the study by [1]. The pre-processing is divided into mining, data clustering, and simulated annealing.
two stages where the first stage is applied to magnetogram and In the next section, different potential tracking techniques will
continuum images and the second stage to correlate between be discussed.
these images. Stage 1 uses a filtering process [5] to remove
IV. T RACKING TECHNIQUES
any irrelevant noise, detect the solar disk, determine its radius
and center, and to create the mask that is used in the filtering A. Scale-invariant feature transform
process. The solar coordinates are calculated for the image, The Scale-Invariant Feature Transform (SIFT) is an algo-
such as the position angle, latitude and heliographic longitude rithm that transforms an image into a large collection of
by implementing various equations. local feature vectors. Each local feature vector is invariant to
The second stage only uses magnetogram images to corre- numerous features, such as image scaling, partially invariant
late both magnetogram and Michelson Doppler Images (MDI) to illumination changes, 3D projection or affine, rotation and
from the first stage. Using magnetogram and continuum im- translation. This technique is based on the behaviour of the
ages has its drawbacks in terms of the time it was captured cerebral cortex’s complex cells of mammalian vision model.

Authorized licensed use limited to: NED UNIV OF ENGINEERING AND TECHNOLOGY. Downloaded on August 07,2024 at 06:07:18 UTC from IEEE Xplore. Restrictions apply.
 1. Figure 3 presents the obtained longitude positions from
the initial tracking sunspot on the x-axis with the longitude
positions expected form the differential rotation on the y-axis.
In equation 1 the θ refers to the latitude, with coefficients
A = 2.6662, B = −0.3441, C = −0.5037 that was used in
this study. Equation 1 works on the computation of the initial
observed longitude, from the first detection of the chosen
sunspot.

Ω = A + B sin2 θ + C sin4 θ (1)

Fig. 2. SIFT technique The A coefficient refers to the rotation rate at the solar
equator. The coefficients B and C refers to partial linear
dependence between them which also increases the derived
A staged filtering approach is used to identify the scale- variance in B and C [8].
invariant features. A Difference-of-Gaussian function is used
as the first step in the SIFT technique where its maxima
or minima is identified in key locations of scale space. The
sampled local image region is described by the points that
generate a feature vector where the sample is relative to
its scale-space coordinate frame. By blurring image gradient
locations, partial invariance is achieved on local variations
such as 3D projections or affine. The feature vectors achieved
are then known as SIFT keys. SIFT keys are derived through
indexing and the nearest neighbour approach, after candidate
object models were identified. The presence of the object is
indicated when at least three SIFT keys are present and is the
same on the model parameters with low residual. The use of
SIFT keys may cause substantial levels of occlusion due to
having multiple SIFT keys but can still retain high levels of
reliability in the image. SIFT is relevant to this study because
it can recognize perspective projection up to twenty degrees
Fig. 3. Correlation between computed and measured longitudes
of the rotation of a 3D object, in this study it is the Sun.
The SIFT technique attempts to identify locations in the
In Figure 4 the initial image of the chosen sunspot which
image scale space, that are invariant with respect to being
will be tracked is shown. After the initial continuum image
minimally affected by noise, small distortions, image scaling,
(left) is chosen, the sunspot’s longitude positions are obtained
translation and rotation [6]. Key locations are selected at the
and plotted across the solar disk (right) from tracking the
minima and maxima of a Difference-of-Gaussian function
sunspot.
applied in the scale space to achieve a high level of efficiency
Another method exists called the local correlation tracking
and rotation invariance. By using re-sampling between each
(LCT) technique that is the most common applied technique in
level of an image pyramid, the computation of this process
the tracking of sunspots [9]. The LCT technique was employed
becomes very efficiently. This process characterization of an
the first time by [10] and then again by [11] and is used for
image is also stable since SIFT locates key points at scales
the calculation of proper motions of the solar granulation. The
and regions of high variation.
LCT technique is based on two consecutive images, that is
used for finding the best cross-correlation between the two
images. In a study related to the LCT technique [12], it became
B. Existing system that use sunspot tracking known that the LCT technique can produce some errors. The
The method used for the tracking of sunspots works in errors are caused by the method itself when it measures the
coherence with the type of detection methods used for iden- intensity changes. These intensity changes may reflect phase
tifying sunspots [7]. The tracking of sunspots is a simple velocities and may also be related to plasma motions.
task in theory where it can be summarized as following the Other errors that may be produced by the LCT, are through
centroid (centre of gravity) of a specific detected sunspot the phenomena known as the shrinking-sun effect that is
across consecutive images. caused by stationary flows [13]. However, artefacts exist that
The program that was created by [7] is an automated can fix it by subtracting the flow maps from the time average
tracking system that plots out the longitude values of a of the velocities from each other. These artefacts used in
specific sunspot with asterisks (Figure 3) based on equation correcting the errors found in the LCT technique have provided

Authorized licensed use limited to: NED UNIV OF ENGINEERING AND TECHNOLOGY. Downloaded on August 07,2024 at 06:07:18 UTC from IEEE Xplore. Restrictions apply.
 data on worst case scenarios of 20-30% in underestimations bounding boxes are then passed to the object tracker, which
of velocities [14]. then tracks these sunspots over the span of the video clip. In
Figure 5 the first five phases of the artefact are displayed, i.e.
(1) the initial image, (2) the Canny Edge Detection that is
applied to the original image, (3) the bounding boxes that are
placed on the contours of the edges, (4) the bounding boxes
are then placed on the original image, and (5) the final result
where the bounding boxes are merged that are closest to each
other.
An appropriate time interval had to be identified to create a
dataset consisting of useful continuum images where sunspots
occurred. February 2014 was identified as the best time due
to it being in the middle of solar cycle and its peak time
Fig. 4. Sunspot tracking example for sunspots to be found. After the date was chosen, the data
was downloaded in the .fits file format due to it being the
most commonly used digital file format in astronomy. The .fits
standard is useful for it contained special features for scientific
data, such as provisions for describing photometric and spatial
V. S YSTEM DESIGN calibration information, together with image origin metadata.
Once the dataset was finished downloading the intensity im-
It was identified that sunspots could pose a threat to various ages were extracted from the .fits files and converted to the
types of technology, such as communication, power stations .png file format. The reason for using .png file format is due
and aviation. Thus, an artefact was developed to detect and do it being a lossless format, meaning that pixel information
track sunspots on images of the Sun that will be used to is not lost. After all the images were extracted from the .fits
help with space forecasting of coronal mass ejections through files, a video was compiled from the dataset that were used
sunspots. An object detection and tracking system was created for tracking purposes.
for space weather forecasting. The sunspots were identified The identification process is contained in the second phase
on continuum images as well as tracked over its lifecycle. of the artefact. This phase was key to ensure that the following
The artefact consisted out of two parts, both of which were steps were executed correctly due to them being based on this
developed using the Python programming language. step’s results. In the identification process, most functions in
The artefact made use of a dataset consisting of continuum the artefact and starts with a function that draws the initial
images, also known as intensity images and the dataset was bounding boxes. The key features in this function was based
created by downloading images that were captured over five on the Canny Edge Detection method and the finding of the
days. These images were then combined to form a video clip. contours to draw the bounding boxes on.
The use of the video was due to natural phenomena where the To use these methods OpenCV was chosen as it provided
Sun is currently in a quiet time and thus, none or few sunspots efficient methods to help with the development of the artefact.
are currently emerging. The video provided sufficient test data OpenCV is an open source machine learning and computer
of when the Sun was in its peak of the solar cycle, resulting in vision software library that contains more than 2500 optimized
more frequent sunspots to test the artefact with. The artefact algorithms. These algorithms include identifying objects, de-
uses the first frame of the selected video clip to apply the tecting and recognizing faces, tracking moving objects and
image detection technique on after which the initial detected much more. OpenCV has various interfaces such Java, C++,
sunspots bounding boxes are stored. The coordinates of these MATLAB and Python and supports operating systems such as
Android, Windows, Mac OS and Linux. For this study openCV
was used using Python in a Linux environment.
The Canny edge detection is a well known and popular edge
detection algorithm that was developed by John F. Canny [15].
The steps of the algorithm includes noise reduction, finding the
intensity gradient of the image, non-maximum suppression and
hysteresis thresholding. The noise reduction in the algorithm
is needed due to edge detection being susceptible to noise that
is present in the image. Thus, a 5x5 Gaussian filter is applied
to remove noise from the image. In Figure 6 an example of a
Gaussian mask is given that can be applied to any image for
noise reduction. The next step is to find the intensity gradient
of the image where the smoothed image is filtered with a Sobel
Fig. 5. Phases of the artefact kernel. The Sobel kernel is applied in both vertical(Gx ) and

Authorized licensed use limited to: NED UNIV OF ENGINEERING AND TECHNOLOGY. Downloaded on August 07,2024 at 06:07:18 UTC from IEEE Xplore. Restrictions apply.
 horizontal(Gy ) direction in order to get the first derivative in Once the identification phase was completed, the final phase
those directions. was conducted. In this phase the tracking of the sunspots
was done by using another openCV method known as ”Mul-
tiTracker”. The reason for implementing tracking into the
artefact from a technical perspective was due to tracking
being more efficient than detection on each frame. Once the
initial detection was done, a lot about the objects’ appearances
are known in the previous frame. Not only is the previous
appearance of the object known, but so is its coordinates
or location on the image. By processing this knowledge, the
tracker is able to predict the next location of the objects. This
method of tracking is effective, but is often more efficient
when also performing the detection on every nth frame to
ensure that the object is still being tracked correctly. Although
Fig. 6. Canny 5x5 filter
tracking is faster than detection, tracking is more prone to
errors in its process. This is where the detection method is
often run again to ensure the tracking algorithm has not made
q an error.
EdgeGradient(G) = G2x + G2y , (2) The tracking section started with a list of coordinates of
the detected bounding boxes being sent to a Python function.
and Once the list is received a function runs through the length
Gy
Angle(θ) = tan−1 ( ). (3) of the list, creating a tracker for each of the bounding boxes.
Gx The default tracker type that was used was the Discriminative
Once the direction and the gradient magnitude is acquired, Correlation Filter with Channel and Spatial Reliability (DCF-
the unwanted pixels are removed which may not constitute CSR) tracker. Kernelized Correlation Filters (KCF) was the
to an edge by performing a full scan on the image. Finally other tracker that was also used for tracking in this study. Once
the threshold is applied which determines whether candidate a tracker was created for each of the boxes, the function moves
edge pixels really are edges and which are not. This is done by into its final phase. The function then plays the video clip
applying two threshold values, one for the minimum value and with every frame being scanned, redrawing the new boxes and
one for the maximum value, such as in Figure 7. Any edges sending the updated location of the objects in the subsequent
that has an intensity value below the minimum threshold value frames. A smaller distance threshold represents more boxes
are considered not to be edges and those that are above the being left unmerged and a higher distance threshold results in
maximum value are sure to be edges. The edges found with an fewer boxes and bigger. The last phase was tested with two
intensity value between the maximum and minimum value are video clips consisting of continuum images. The effectiveness
only considered to be edges, depending on their connectivity. of the tracking was influenced mainly by the second phase
which gave the boxes that needed to be tracked.
VI. R ESULTS
A. Results of identification
The error percentage of the identification phase was mostly
determined by the size of the threshold for the bounding
boxes to be merged. If the threshold was larger, fewer boxes
were displayed and a smaller threshold resulted in more
boxes. With a larger threshold, the artefact was more prone
to miss identification and false positives of sunspots, whereas
Fig. 7. Canny maximum and minimum threshold the smaller threshold detected all the sunspots fewer false
positives.
For the purpose of this study, a threshold of 100 pixels was
The Canny technique was applied to the initial image in used as the standard for measuring the error percentage of
order to find the edges of the sunspots. Once the sunspots were the identification phase. The error percentage is calculated by
found, the openCV technique ”findContours” was applied to equation 4:
place the bounding boxes on the edges. The distance of each
centre point of the boxes were then computed in order to find experimental value − accepted value
|error| = × 100%. (4)
boxes that are within a certain distance from each other. If the accepted value
boxes were in the specified range, they were merged together A range of ten continuum images were used each with a
and a new bigger bounding box was formed. threshold of a 100 pixels between each bounding box centre

Authorized licensed use limited to: NED UNIV OF ENGINEERING AND TECHNOLOGY. Downloaded on August 07,2024 at 06:07:18 UTC from IEEE Xplore. Restrictions apply.
 point. The following was the percentage error results of each VIII. C ONCLUSION
image: Space weather plays a major role in keeping these systems
1) Eight out of eleven sunspots were detected resulting in fully functioning and does so by executing numerous tasks
27% error percentage. that involves the Sun and its sunspots. There exist several
2) Six out of six sunspots were detected resulting in 0% systems that can detect, track and classify sunspots, but not
error percentage. all systems use the same techniques to achieve their maximum
3) Nine out of ten sunspots were detected resulting in 10% accuracy. An artefact was developed to test if it was possible to
error percentage. develop a system that can remove manual labour and replace it
4) Nine out of nine sunspots were detected resulting in 0% with an automated system. The result was a working proof of
error percentage. concept that has potential in future work. Potential future work
5) Twelve out of thirteen sunspots were detected resulting would include a system that implements other identification
in 8% error percentage. techniques to increase the accuracy and to successfully track
6) Seven out of seven sunspots were detected resulting in sunspots that disappear and reappear on the Sun.
0% error percentage.
R EFERENCES
7) Ten out of twelve sunspots were detected resulting in
17% error percentage. [1] T. Colak and R. Qahwaji, “Automatic sunspot classification for real-
time forecasting of solar activities,” Proceedings of the 3rd International
8) Eight out of eleven sunspots were detected resulting in Conference on Recent Advances in Space Technologies, RAST 2007, pp.
27% error percentage. 733–738, 07 2007.
9) Eleven out of eleven sunspots were detected resulting in [2] A. Hanslmeier, The Solar Atmosphere and Active Regions.
Dordrecht: Springer Netherlands, 2007, pp. 47–95. [Online]. Available:
0% error percentage. https://doi.org/10.1007/978-1-4020-5604-8 3
10) Four out of five sunspots were detected resulting in 25% [3] J. J. Curto, M. Blanca, and E. Martı́nez, “Automatic sunspots detection
error percentage. on full-disk solar images using mathematical morphology,” Solar
Physics, vol. 250, no. 2, pp. 411–429, Aug 2008. [Online]. Available:
This resulted in an average error percentage of 11.6%. https://doi.org/10.1007/s11207-008-9224-6
[4] Carley, Eoin P., Vilmer, Nicole, Simões, Paulo J. A., and Ó Fearraigh,
Brı́an, “Estimation of a coronal mass ejection magnetic field strength
B. Results of tracking using radio observations of gyrosynchrotron radiation,” Astronomy
and Astrophysics, vol. 608, p. A137, 2017. [Online]. Available:
The tracking of the sunspots was done on only two videos. https://doi.org/10.1051/0004-6361/201731368
The distance threshold once again determined their sensitivity [5] R. Qahwaji and T. Colak, “Automatic detection and verification of solar
features,” International Journal of Imaging Systems and Technology,
and accuracy in tracking. A standard threshold of a hundred vol. 15, no. 4, pp. 199–210, 2005.
pixels were used again for measuring purposes. Two types of [6] D. G. Lowe, “Object recognition from local scale-invariant features,”
tracking were used, DCF-CSR and KCF that was discussed in 1999, pp. 1150–1157 vol.2.
[7] S. Goel and S. K. Mathew, “Automated detection, characterization,
previous section. The DCF-CSR was the most accurate tracker and tracking of sunspots from SoHO/MDI continuum images,” Solar
that resulted in 0% error percentage whereas the KCF resulted Physics, vol. 289, no. 4, pp. 1413–1431, 2014.
in over 30% error percentage. The reason for the first tracker [8] H. B. Snodgrass and R. K. Ulrich, “Rotation of doppler features in the
solar photosphere,” The Astrophysical Journal, vol. 351, pp. 309–316,
being so successful was due to the DCF-CSR prioritizing 1990.
accuracy over frames per second output. The KCF prioritized [9] Campos Rozo, J. I., Utz, D., Vargas Domı́nguez, S., Veronig, A.,
frames per second output over accuracy which gave it such and Van Doorsselaere, T., “Photospheric plasma and magnetic field
dynamics during the formation of solar AR 11190,” Astronomy
high error percentage. The only problem in the tracking was and Astrophysics, vol. 622, p. A168, 2019. [Online]. Available:
where the sunspots, that was being tracked, were to suddenly https://doi.org/10.1051/0004-6361/201832760
shrink and disappear over the video. This resulted in the [10] L. J. November, G. W. Simon, T. D. Tarbell, and A. M. Title, “Precise
proper motion measurement of solar granulation,” vol. 18, p. 665, Mar
trackers not knowing what to track, and causes them to track 1986.
random areas on the Sun. [11] L. J. November and G. W. Simon, “Precise proper-motion measurement
of solar granulation,” The Astrophysical Journal, vol. 333, p. 442, Oct
1988.
VII. D ISCUSSION [12] T. Roudier, M. Rieutord, J. M. Malherbe, and J. Vigneau, “Determination
of horizontal velocity fields at the sun’s surface with high spatial and
In the identification phase, if the boxes that were identified temporal resolution,” Astronomy and Astrophysics, vol. 349, pp. 301–
were too large and the sunspots in them were very small, the 311, Sep 1999.
[13] J. Lisle and J. Toomre, “Cause of shrinking sun effect in local correlation
trackers would incorrectly track the sunspots. The major cause tracking and impacts on the mapping of SSW flows,” in SOHO 14 Helio-
of the misidentification of sunspots were due to them being and Asteroseismology: Towards a Golden Future, ser. ESA Special
slightly visible to the naked eye, but the Canny edge detection Publication, D. Danesy, Ed., vol. 559, Oct 2004, p. 556.
[14] A. Asensio Ramos, I. S. Requerey, and N. Vitas, “DeepVel: Deep
method threshold was unable to detect them. Additional fea- learning for the estimation of horizontal velocities at the solar surface,”
tures could be added such as, adding classification of sunspots Astronomy and Astrophysics, vol. 604, p. A11, Jul 2017.
to classify the size and danger of a sunspot into a category. [15] J. Canny, “A computational approach to edge detection,” IEEE Transac-
tions on pattern analysis and machine intelligence, vol. PAMI-8, no. 6,
Neural networks could be added to help with the classification pp. 679–698, 1986.
as well as to use different frequencies of images such as the
magnetogram spectrum.

Authorized licensed use limited to: NED UNIV OF ENGINEERING AND TECHNOLOGY. Downloaded on August 07,2024 at 06:07:18 UTC from IEEE Xplore. Restrictions apply.