Aurora is a very important geophysical phenomenon in the high
latitudes of Arctic and Antarctic regions, and it is important to make a
comparative study of the auroral morphology between the two hemispheres.
Based on the morphological characteristics of the four labeled dayside
discrete auroral types (auroral arc, drapery corona, radial corona
and hot-spot aurora) on the 8001 dayside auroral images at the Chinese Arctic Yellow
River Station in 2003, and by extracting the local binary pattern (LBP)
features and using a
Aurora is caused by the collision of charged solar wind particles with the neutral particles in the polar atmosphere. Study of the morphology and evolution of aurora is not only helpful for revealing the solar wind–magnetosphere–ionosphere coupling processes and their internal mechanisms but also provides important physical principles for space weather forecasting (Nishimura et al., 2010; Hu et al., 2009, 2010, 2012, 2013, 2017a, b; Han et al., 2015, 2016, 2017).
As auroral research continues, comprehensive auroral observation has become
an important field of polar scientific research for various countries in the
world. All-sky cameras (ASCs), which have high spatial resolution and can
make broad views and long-term continuous observations, are one of the most
effective tools for auroral observation and the investigation of
magnetospheric structure and dynamic processes. Therefore, there are many
ground-based auroral observatories in the Arctic and Antarctic (Yang et al., 2000; Ebihara et al.,
2007; Hu et al., 2009, 2017c). With the establishment of many ground-based
stations, those optical instruments produce hundreds of millions of images
annually (Syrjäsuo and Partamies, 2012), leading to an urgent need for automatic auroral image
analysis techniques (Syrjäsuo et al., 2001, 2004, 2007).
Up to now, many studies have been conducted using data from only one station.
Syrjäsuo and Donovan (2004) first explored computer vision techniques to
auroral image classification. They analyzed 350 000 Gillam ASC images from the
CANOPUS (Canadian Auroral Network for the OPEN Program Unified Study)
program during 1993–1998 and categorized them into no aurora, arcs, patchy
aurora, omega-band, and other auroral structures. Based on the ASC data
observed in Kilpisjärvi, Syrjäsuo and Partamies (2012) created an
automatic detection scheme for aurora (i.e., whether aurora existed or not). Ebihara et al. (2007)
analyzed the quasi-stationary auroral patches based on the ASC images
observed between
However, each of the aforementioned studies was performed on one station;
although the MIRACLE network (Syrjäsuo et al., 1998) includes several
stations, the data were studied as a whole (Rao et al., 2014; Savolainen et
al., 2016). Recently, Pulkkinen et al. (2011) reported auroral occurrence
by using auroral observations from five stations in Fennoscandia and Svalbard
in 2000–2009, and Partamies et al. (2014) used 1 million
auroral images captured at five camera stations in Finnish and Swedish
Lapland in 1996–2007 to study the solar cycle and diurnal
dependence of auroral structures; the data in different stations were still
analyzed together. Based on the synoptic distribution of the average auroral
intensity in the Arctic and Antarctica, Hu et al. (2014) discussed the
hemispheric asymmetry of the dayside auroral oval structures, yet they did
not consider auroral morphology because they were limited to manual analysis. At present,
there are very few comparative studies based on multiple stations, especially
with regard to contrastive study of auroral morphology between the Northern and Southern
Hemisphere. The different dynamic processes in the magnetosphere can result
in different morphological characteristics of the aurora. By comparing the
morphological characteristics of the auroras between the Northern and
Southern Hemisphere, an investigator can study the difference or similarity
in ionospheric responses between the Northern and Southern Hemisphere which
result from the dynamic processes in the magnetosphere. In this paper, the
LBP descriptor is exploited to characterize the ASC images, and the
The remainder of this paper is organized as follows. In Sect. 2, the data and LBP-based representation method are introduced. The automatic recognition experimental results of dayside aurora in SPS and YRS are presented in Sect. 3. Section 4 is the discussion. Finally, the conclusion is drawn in Sect. 5.
The auroral data explored in this paper were captured by the all-sky cameras
at two stations. (1) The South Pole Station (SPS) in Antarctica is
located at 90.0
The optical instruments at SPS can make 24 h surveys of auroral emissions with a temporal resolution from a few seconds to dozens of seconds during the winter season from April to August. In this paper, we focus on the dayside aurora (09:00–22:00 UT/05:24–18:24 MLT) from May to August to avoid the influence of daylight. Optical instruments cannot capture clear auroral images in cloudy or foggy weather, and after eliminating those invalid data captured under poor weather conditions or without the information of aurora, altogether 211 days of auroral observations are selected from May 2003 to August 2005 to constitute the SPS data set, referred to as auroral test dataset 1 (ATD1), which consists of 39 335 images.
Calculation process of the basic LBP operator.
Prior to analyzing these ASC images, some preprocessing steps are performed:
Rescaling. Few pixels have an intensity greater than 8000; therefore
all images are stretched with a cutoff value of 8000. Image stretching can
preserve pixels' relative intensity and enhance image contrast, making
auroral images easier to classify. Cropping and rotation. We first
crop the images to make the central field of view the center of the ASC
images; secondly, the images are rotated counterclockwise by
125.57 Orientation adjustment is done in the direction of east and west in order to keep images consistent with those of YRS (left: east; right: west).
The ASCs at YRS continue to produce auroral images on three wavelengths (427.8, 557.7 and 630.0 nm) for 24 h per day with a temporal resolution of 10 s during the whole winter season (October to March). In this paper, we concentrate on the dayside aurora (03:00–15:00 UT/06:00–18:00 MLT) at 557.7 nm from November to February. After removing those uninformative images captured under poor weather conditions, altogether 249 days of auroral data are selected from December 2003 to January 2009 to constitute the YRS data set. It is divided into two parts: (1) the auroral train dataset (ARD) consists of 8001 images captured from December 2003 to January 2004, and (2) the auroral test dataset 2 (ATD2) includes 65 361 images generated from October 2004 to February 2009.
Some preprocessing steps are also conducted to YRS images.
Removing system noise. System noise in ASC instruments is caused by dark
current;
therefore all images are subtracted from the image of dark current. Masking
and cropping. In order to have the same size with the ASC images at SPS, a
circle mask with a radius of 199 pixels is used to cut off the outer regions
and the image size is cropped to 398 Rotation. The
images are rotated counterclockwise by 61.1 Rescaling. All
images are stretched with a cutoff value of 4000. Note that although the Rayleigh
intensity is different between YRS images and SPS images, it does not
influence the following results since we extract the texture feature of ASC
images as follows.
The original LBP operator was introduced by Ojala et al. (1996) primarily for texture classification, and we have proved that it has excellent ability to characterize the spatial texture of auroral images (Wang et al., 2010; Yang et al., 2012). In this paper, the improved LBP operators (Ojala et al., 2002) with a partition scheme are applied to represent ASC images.
LBP is a simple and efficient texture descriptor; it characterizes a local
region by comparing the relative gray values between the central pixel and
its neighboring pixels. Figure 1 shows the calculation process of the basic
LBP operator. The central pixel value is 100, and its 3
In order to capture the dominant features with large-scale auroral
structures, we use the improved LBP (Ojala et al., 2002; Wang et al., 2010),
which is based on circular symmetric neighborhoods of different sizes, to
represent the auroral images. Specifically, the parameters of sampling points
The LBP extraction process of the auroral image. The image illustrated was captured at 05:31:51 on 21 December 2003.
According to the characteristics of auroral images, each image is divided into 3 rows and 6 columns, obtaining 18 rectangular blocks (Fig. 2a), and the histogram of the LBP patterns is calculated in each block (Fig. 2b). Finally, these individual histograms are concatenated into a global descriptor for ASC images (Fig. 2c).
In this paper, the auroral images are classified into arc, drapery corona, radial corona, and hot spot based on the spectral and morphological signatures (Hu et al., 2009). (1) Arc aurora is a striped auroral form with east–west-extending and narrow north–south-spanning characteristics, and often multiple auroral arcs appear simultaneously in the sky. (2) Drapery corona is a weak display with the features of east–west elongated bands, and often multiple parallel rays appear at the same time. (3) Radial corona has clear radial-like structures which spread from the zenith in all directions. (4) Hot-spot aurora has complex structures, showing ray-like coronal auroras, irregular patches of auroral intensity enhancement (vortex, spots) and arc-like auroral mixed morphologies. Previous investigations (Hu et al., 2009; Wang et al., 2010) reveal that these auroras mainly appear in the hot-spot auroral active region, so they are referred to as hot-spot auroras. More detailed descriptions of the four auroral types can be in Hu et al. (2009) and Wang et al. (2010). In addition, there are some images that the aurora morphology has difficulty classifying into the four auroral types. We also refer to these as the “unknown” type.
Based on the above descriptions, and referring to the synchronous ASC images
at wavelengths of 427.8 and 630.0 nm (Wang et al., 2010), ASC images in the ARD are manually labeled according to the abovementioned four categories. In order
to avoid very similar morphology between adjacent ASC images (because of
the short sampling interval of 10 s at YRS), the ARD is constructed
by extending the interval between adjacent images to
Content-based image retrieval experiments are performed on ARD and ATD1
to examine the morphology difference between auroral images at YRS and SPS.
The retrieval results are visually estimated to see whether each retrieved image
has similar auroral morphology to its query image. Chi-square (
Comparison of the four auroral types at SPS and YRS by retrieval experiments. The images from labeled ARD of YRS are used as query images, and the images of SPS are the retrieval results from ATD1 which are the most similar to each query image.
Based on the spectral signatures of dayside auroras, Hu et al. (2009) partitioned the dayside oval into four auroral active regions, i.e., a pre-noon warm spot (“W”, 07:30–09:30 MLT), a midday gap (“M”, 09:30–13:00 MLT), a post-noon hot spot (“H”, 13:00–15:30 MLT), and the dusk aurora sector (“D”, 15:30–17:00 MLT). In this section, supervised classification experiments are conducted on ARD to estimate the classification accuracy in different auroral active regions.
Specifically, the widely used
Ten-fold cross-validation experiments are performed on the ARD. Images of
the four auroral types are separately divided into 10 parts randomly, of
which 9 parts are used to train the classifier, and the remaining part is
used as the testing set to evaluate the classification effectiveness. The
training–testing ratios in each auroral active region are not strictly
guaranteed to be
The number of images belonging to four auroral types in different auroral active regions, including the number of the testing set in 10-fold cross-validation experiment and all of the images in the ARD (in parentheses).
Experiment performance
from 100 iterations (mean accuracy
From Table 2, we can conclude the following. (1) The NN classifier works best. Two adjacent images in the ARD are picked out in a short time interval; specifically, there are 1690 out of 8001 images captured with intervals less than 2 min (Yang et al., 2012). ASC images with such short intervals always have similar morphology; therefore the NN classifier works best in the experiments. (2) In the W region, the classification accuracy changes considerably and has a large standard deviation value. This is because there are very few hot-spot auroras (fewer than 4 % of the total number), as shown in Table 1; the accuracy will change considerably even if only one image is categorized into different types. (3) In the D region, the classification accuracy also changes substantially and has a large standard deviation value too. The reason for this is also presented in Table 1; most images in the D region are arc auroras, and very few images belong to other auroral types (fewer than 8 % of arc auroras). Very few images classified into different categories may lead to a large difference in the accuracy of drapery corona, radial corona and hot-spot auroras and with large deviation values. (4) On the whole, the more the image data, the smaller the deviation value. (5) Except for invalid results (standard deviation is greater than 10 %, which is shown by gray shading in Table 2), the accuracies of arc and hot-spot auroras show a significant decrease at the M region, while drapery and radial coronas have a lower accuracy at the H region. (6) The proposed method achieves very good performance: almost all the classification accuracies are higher than 90 %.
In this part, the labeled ARD is used as the training set, and
by exploiting the class information contained in it, we recognize all the
images in ATD1 and ATD2 with a
Unlike ARD, when constructing ATD1 and ATD2, all auroral images are picked
except for the images captured under poor weather conditions or which have no
aurora information; therefore, a rejection classification is needed. The
Classification results for ATD1 of SPS and ATD2 of YRS using 3-NN and 25-NN classifiers.
From Table 3 we can see that (1) there is little difference between 3-NN and 25-NN classifiers, especially for radial corona, arc, and hot-spot auroras, whose differences are less than 1.5 %; (2) the most different auroral type between the two stations is drapery corona, because the texture of the drapery aurora is complex and error-prone; (3) the occurrence percentages of the four auroral types at both stations in the Northern and Southern Hemisphere are very close, of which the arc aurora is about 39 %, while drapery corona is around 33–38 %, radial corona is about 20 %, and hot spot is around 3 %; and (4) the number of “unknown” type obtained by the 25-NN classifier is less than for the 3-NN, because it is more difficult for 3-NN classifiers to get a majority vote.
The temporal distributions of the four auroral categories are presented in Fig. 4. For ATD1 and ATD2, the temporal axis is divided into 39 and 36 bins, respectively, with 20 min durations for better display. The number of images within each bin for each category is first counted, and by normalizing the total number of images in the same bin, the occurrence distributions are obtained. There are no significant differences between the distributions obtained by 3-NN and 25-NN, so we only show that of the 25-NN in Fig. 4. Panels (a) and (b) of Fig. 4 shows the distribution of all images in ATD1 and ATD2.
Temporal occurrence distributions of auroral types at SPS
From the second panel to the bottom panel, Fig. 4 shows the occurrence distribution of the four auroral categories and the unknown type. At the very top of Fig. 4, four active regions proposed by Hu et al. (2009) are distinguished by bold dashed lines, while the two states of the midday gap are partitioned by a thin dashed line. The four auroral types dominate different dayside oval regions The peaks of four auroral types each fall into different active regions.
From Fig. 4, we can see that the occurrence distributions at both stations are similar. (1) The occurrence distributions of arc auroras show a distinct asymmetric double peak between pre-noon and post-noon (Akasofu and Kan, 1980; Liou et al., 1997; Meng and Lundin, 1986; Newell et al., 1996; Hu et al., 2009). (2) Two corona auroras, drapery and radial, have similar auroral morphologies, and both predominantly occur before 13:00 MLT. (3) The hot-spot auroras mostly occur in region H and have a distinct small peak around 13:30 MLT. (4) There are a few atypical images in ATD1 and ATD2 that are classified as unknown. (5) In addition to the similar distribution trends of the four auroral categories at both stations, the occurrence ratios of each category in each MLT at both stations are also very close.
Although there are so many similarities, differences also exist in the auroral occurrence distributions of the two stations. (1) Drapery and radial coronas show different peak positions at the two stations. (2) Although the hot-spot aurora has an occurrence peak around 13:30 MLT at both SPS and YRS, the duration of YRS is much longer than SPS.
The discrete aurora includes acceleration mainly caused by two physical mechanisms: one is quasi-static electric fields, producing monoenergetic auroral precipitation that always appears in the “inverted V” electron spectrum, while the other is dispersive Alfvén waves, producing broadband auroral electron precipitation (Newell et al., 2009). Ionospheric satellite observations show that monoenergetic precipitation mostly exists in the 14:00–19:00 MLT region of the post-noon auroral oval, followed by the 06:00–09:00 MLT region of the pre-noon auroral oval (the incidence and electron precipitation energy flux of the latter are lower than the former), and the noon (09:00–14:00 MLT) region is the least likely region for observing monoenergetic electron acceleration (Newell et al., 1996, 2009). Such a distribution is consistent with that of the arc aurora. As shown in Fig. 4, at both stations, the occurrence percentages of arc aurora at the 14:00–18:00 MLT region of the post-noon oval are greater than 50 %, and the percentages at the 06:00–09:00 MLT region of the pre-noon oval are 30–50 %, while they are less than 30 % at 09:00–14:00 MLT of the noon region, especially around 12:00 MLT of the noon the percentages are even less than 10 %. The similar occurrence distributions between the two stations demonstrate that the arc auroras with electron spectrum characteristics of “inverted V” structures are closely associated with the quasi-static electric field acceleration.
Dayside coronas (drapery, radial and hot spot) have distinct filament structures, which indicates that (1) the corona aurora is excited in a broad altitude range and (2) the energy distribution of these precipitating electrons that excite coronas is very broad. (The excitation altitude of aurora is closely related to the energy of precipitating electrons: the higher the energy, the lower the altitude of precipitating electrons entering into the ionosphere.) Satellite spectrum probes have proved that the precipitating electron spectrum of dayside coronas has the signatures of the broadband auroral electron precipitation (Hu et al., 2009). Ionospheric satellite observations show that the broadband electron precipitation in the dayside auroral oval mostly occurs at the 06:00–15:00 MLT region, and it occurs more often at pre-noon than post-noon (Newell et al., 2009). The occurrence rate of dayside corona auroras at both stations in the Northern and Southern Hemisphere also dominates the 06:00–15:00 MLT regions, which is similar to that of the satellite detection. In addition, satellite observations obtained at higher altitudes (> 6000 km) show that dispersive Alfvén waves primarily occur at the 06:00–15:00 MLT region of the dayside auroral oval (Chaston et al., 2007). Therefore, dayside corona auroras are closely related to dispersive Alfvén waves. However, the ground-based optical observations demonstrate that the dayside corona auroras can be classified into three types: drapery corona auroras, radial corona auroras, and hot-spot auroras. These corona auroral types are possibly related to the propagation process of dispersive Alfvén waves at different magnetosphere boundary layers, which results in the difference of the three corona auroral types between the two stations. Such an inference needs further confirmation by combined analysis of satellite and ground-based optical observations.
Based on previous studies of the morphological classification of dayside auroras (Hu et al., 2009), and by applying image processing and pattern recognition techniques on the ASC observations at SPS (during years 2004–2006) and YRS (during years 2003–2009), this paper created an automatic recognition scheme for dayside auroral morphology in the Southern and Northern Hemisphere and a statistical analysis of the distribution of dayside auroral types. Experimental results show that in both the Southern and Northern Hemisphere, the dayside arc auroras primarily occur at post-noon (14:00–18:00 MLT) and pre-noon (06:00–09:00 MLT) regions but mostly occur at post-noon, while between the two peaks, the noon region (09:00–14:00 MLT) forms a “midday gap”. Dayside corona auroras mostly occur at 06:00–15:00 MLT regions. The distribution of arc and corona respectively corresponds to the occurrence rate of quasi-static electric field acceleration and dispersive Alfvén wave acceleration on the dayside auroral oval.
All data in the paper are public, and please connect the
first and corresponding authors for the data. Data were issued by the
Data-sharing Platform of Polar Science (
The authors declare that they have no conflict of interest.
We acknowledge Yusuke Ebihara at Kyoto University for providing the ASC
auroral observation data of the South Pole Station
(