The auroral oval boundary represents an important physical process with implications for the ionosphere and magnetosphere. An automatic auroral oval boundary prediction method based on deep learning in this paper is applied to study the variation of the auroral oval boundary associated with different space physical parameters. We construct an auroral oval boundary dataset to train our proposed model, which consists of 184 416 auroral oval boundary points extracted from 3842 images captured by the Ultraviolet Imager (UVI) of the Polar satellite and its corresponding 18 space physical parameters selected from the OMNI dataset from December 1996 to March 1997. Furthermore, several statistical experiments and correlation analysis experiments are performed based on our dataset to explore the relationship between space physical parameters and the location of the auroral oval boundary. The experiment results show that the prediction model based on the deep learning method can estimate the auroral oval boundary efficiently, and different space physical parameters have different effects on the auroral oval boundary, especially the interplanetary magnetic field (IMF), geomagnetic indexes, and solar wind parameters.

An auroral oval is a circular belt of auroral emission around magnetic poles (Loomis, 1890; Akasofu, 1964). The auroral oval poleward and equatorward boundaries are related to geophysical parameters, which can indicate for the coupling process among the solar wind, ionosphere, and magnetosphere, for example, the polar cap ionosphere, which is considered an area of the opening magnetic field inside the auroral oval poleward boundary. This area is closely related to energetic particle entrance from the heliosphere to the earth's atmosphere. So, the segmentation and prediction for the auroral oval boundary are very significant for studying certain physical events.

In the past few decades, scholars have conducted extensive research on the relationship between the location of the auroral oval boundary and space physical parameters (Niu et al., 2015). In early research, Feldstein proposed that the position of the auroral oval boundary is correlated with the Q index of magnetic activity on the nightside of the earth (Feldstein and Starkov, 1967). Starkov and Holzworth stated that the inner and outer boundaries of the auroral oval can change with geomagnetic indexes and the interplanetary magnetic field (IMF) (Holzworth and Meng, 1975; Holzworth and Meng, 1984; Starkov, 1994a). The conclusions in this paper are based on mathematical statistics. Therefore, Starkov designed some simple formulas to describe the relationships between the specific physical parameter and different types of auroras. Variations of the sizes of the polar cap, auroral oval, and diffuse aurora were regarded as three independent function variables of the AL index (Starkov, 1994b). Since then, many scholars have explored the connections between different physical parameters and the auroral oval boundary or other auroral events. Carbary constructed a Kp-related model of the auroral oval boundary by binning Ultraviolet Imager (UVI) images from different months (Carbary, 2005). To describe the particle precipitation characteristics, Zhang and Paxton proposed a Kp-dependent model of the mean energy and energy flux precipitating electrons in the auroral oval (Zhang and Paxton, 2008). Sigernes et al. used a Kp-based function to calculate the size and position of the auroral oval and compared the Kp-dependent model with methods proposed by Zhang and Starkov to explain the superiority of his proposed model (Sigernes et al., 2011). Milan proposed a model based on average protons and electrons of auroral images from 3 years observed by the IMAGE spacecraft. The experiment demonstrated that Kp, solar wind parameters including solar wind velocity, density, and pressure, IMF magnitude, and orientation have an effect on the intensity and shape of the auroral oval (Milan, 2010). Since 2010, there have been more and more new methods to construct a connection between the position of the auroral oval boundary and space physical parameters with the development of machine learning. Hu and Yang used the segmentation results of the auroral oval obtained from the UVI on the Polar satellite to build a connection between the positions of the auroral oval boundary and the auroral electrojet (AE) index, the IMF, and solar wind parameters by using a multiple regression method (Hu et al., 2017; Yang et al., 2016). Ding presented a C-means clustering algorithm based on fuzzy local information to extract the auroral oval poleward and equatorward boundaries from merged images with filled gaps captured from both GUVI and SSUSI (Ding et al., 2017). However, the position of the auroral oval boundary is not determined by one space physical parameter; those methods mentioned above only used one or several space physical parameters to explore the relationship between space physical parameters and the auroral oval boundary. We cannot determine whether other space physical parameters can influence the location or size of the auroral oval, and we also do not know whether the mapping relationship between space physical parameters and the auroral oval boundary is linear or non-linear.

As we know, machine learning has been applied to many fields, including the medical, traffic, space physics, and other interdisciplinary fields. Recently, deep learning models have led to a series of breakthroughs in image classification, object detection, image recognition, and other fields. Conventional machine learning methods have some limitations for processing complex data, especially in the space physics field. There are no suitable internal features, such as shape and colour. Therefore, many effective machine learning methods cannot obtain satisfactory performance on processing space physics data, while deep learning methods are representation learning methods with multiple levels of representation. They have turned out to be very good at discovering intricate structures in high-dimensional data and multimodal data (LeCun et al., 2015).

In this paper, a new automatic auroral oval boundary prediction model is proposed based on a deep learning method. The experiment results show that the model proposed in this paper can predict the auroral oval boundary accurately by using space physical parameters and the location of the auroral oval boundary at the previous moment. In addition, we explore the effect of every space physical parameter on the auroral oval boundary. The rest of this paper is organized as follows. Section 2 describes our proposed algorithm in detail. The experiment analysis and discussion are given in Sect. 3, including dataset construction, subjective and objective evaluation, the selection of model parameters, and the discussion about the influence of every space physical parameter on the auroral oval boundary. Finally, we draw several conclusions in Sect. 4.

The flowchart of the auroral oval boundary prediction model based on deep learning.

The flowchart of the auroral oval boundary prediction model is shown in Fig. 1.
There are two major steps in our proposed model, pre-training on our dataset
and online prediction. In the training phase, auroral oval images are
usually affected by heavy noise and other interferences. So, the auroral
oval boundary is blurred, and it is difficult to find from the background.
Compared with other image segmentation methods, maximal similarity-based
region merging (MRSM) (Liu et al., 2013) can eliminate the cumbersome
process of adjusting parameters and has better segmentation accuracy. We use
MRSM firstly to extract positions of the auroral oval boundary. The centre of
the auroral oval spatial distribution in the magnetic local time–magnetic latitude
coordinate (MLT–MLAT) is located in the geomagnetic pole. The magnetic
latitude of the auroral oval usually ranged from 57.5 to 73.5

The deep learning network is constructed by a two-layer restricted Boltzmann
machine (RBM) network (Hinton et al., 2006; Yu and Deng, 2011) and a radial
basis function (RBF) network (Łukaszyk, 2004). The computational
processing of the RBM and RBF is illustrated by Eqs. (1)–(4). In the training
phase, the inputs of the RBM network are 18 space physical parameters from the OMNI
dataset and coordinate values of the auroral oval poleward and equatorward
boundaries extracted from segmented UVI images with MRSM. They can be
represented as

The function of the RBF network can make the output of the RBM network infinitely
approximate to the coordinate values of the auroral oval boundary by a radial
basis function. The input of the RBF network is the output of the second layer in
the RBM network. The output of the RBF network is represented as

The auroral oval images used in this paper are captured by the UVI, which is a 2-D snapshot-type camera on the Polar satellite. The UVI
on the Polar satellite acquired more than several million images during
its entire mission. As of April 2008, it is no longer active. There was no
effective observation after 2000, because the Polar satellite changed its view
after 2000. In order to balance the relationship between spatial resolution
and global coverage, the spatial resolution of the UVI is 30 km at apogee, the
charge-coupled device (CCD) array onboard has

The schematic of extracting auroral boundary points.

Space physical parameters selected from the OMNI dataset.

The space physical parameters were downloaded from the NASA OMNI dataset with a different time resolution. It is common knowledge that the IMF, solar wind parameters, and geomagnetic indexes have a time resolution of 1 min, and the other space physical parameters maybe have a higher time resolution. According to the effect derived from other circumstantial factors, such as the time to traverse the magnetosphere and Alfven wave, not all the response times of auroral events are equal to their propagation time. We align the time of all space physical parameters with the time of UVI images in our dataset to avoid the problem of different time resolution between space physical parameters and auroral oval images. In the OMNI dataset, we selected 18 space physical parameters, including the common parameters, which have been verified as being related to the position of the auroral oval boundary (Holzworth and Meng, 1975; Starkov, 1994a, b; Milan et al., 2010; Hu et al., 2017) and some unfamiliar parameters which are never discussed in previous works. Therefore, our dataset includes 184 416 auroral oval boundary points extracted from 3842 UVI images and its corresponding values of 18 space physical parameters. Table 1 shows the 18 space physical parameters which we used in this paper.

In order to evaluate the precision of predicted auroral oval boundary points
by our model, we use the common metric MAE (mean absolute error) to assess
the error between predicted auroral oval boundary points and real auroral
oval boundary points. The MAE can be defined as Eq. (5).

Since the effectiveness of the prediction model is influenced by the number of hidden layer nodes in the RBM network (Hinton, 2012) and the training error of the RBF network, we build two experiments to find the most suitable parameters for our network. For both experiments, space physical parameters and positions of poleward and equatorward boundary points in 24 MLT regions of 3000 UVI images are selected as training samples, and the rest are regarded as test samples. In experiment 1, the training error of the RBF network is set to 4 magnetic latitudes and the numbers of hidden layer nodes in the RBM network are 32, 64, 96, and 128 respectively. We use the average MAE with 100 experiments to verify the stability of our model, because training samples and test samples were divided by a random number. The corresponding MAE is shown in Fig. 3a. From Fig. 3a, MAE reaches the smallest value when the number of hidden layer nodes is set to 32. In experiment 2, the number of hidden layer nodes is set to 32 according to the results in experiment 1. There often has been an overfitting problem when we train a neural network (Krizhevsky et al., 2012). Overfitting can be interpreted as a phenomenon where the model performs well on the training set and unsatisfactorily on the test set. We set different training errors to avoid the overfitting problem. So, the training error of the RBF network is set to 2, 4, 6, and 8 magnetic latitudes empirically. The corresponding MAE is shown in Fig. 3b, and MAE reaches the minimum when the training error of the RBF network is 4 magnetic latitudes. From the two experiment results above, we set the number of hidden layer nodes in the RBM network and the training error of the RBF network to 32 and 4 respectively as the optimal parameters of the deep learning network in the following experiments.

To demonstrate the availability of our proposed model, we compared the
proposed model with a back-propagation (BP) network (Rumelhart, 1986) and
Yang's model (Yang et al., 2016). The subjective prediction results obtained
by the three methods are shown in Fig. 4; circles and squares stand for
poleward boundary points and equatorward boundary points which are obtained
from the segmented image, and “

The different subjective results based on different methods.

The MAE values of different methods.

As we know, the location of the auroral oval boundary is affected by a variety of space physical parameters. Variations of the auroral oval boundary in different MLT sectors are related to different space physical parameters. For the sake of exploring the influence of space physical parameters on the poleward and equatorward boundaries specifically, the boundary points are further processed as follows (Hu et al., 2017). Firstly, all poleward and equatorward boundary points are divided into 24 subsets of poleward and equatorward boundary points according to 24 MLT sectors. Secondly, in every MLT subset, we sort boundary data with respect to the values of all the space physical parameters and divide boundary data into 10 groups evenly. In order to observe the variation tendency of each parameter in different MLT sectors clearly, in every MLT sector, the relationship between each space physical parameter and the location of the auroral oval boundary was represented as a quadratic equation based on the principles of the least square conic fitting (Fitzgibbon et al., 1999). Then, we calculate the locations of poleward and equatorward boundary points for each space physical parameter using this function. Finally, we use the boundary data calculated by the quadratic equation to discuss the influence of space physical parameters on the auroral oval boundary. In this section, we build three statistical experiments to discuss how the IMF, solar wind parameters, and geomagnetic indexes influence the auroral oval boundary, an auroral oval boundary prediction experiment by inputting every single space physical parameter to explore the relationship between the auroral oval boundary and 18 space physical parameters, and a correlation analysis experiment is constructed to study the connection between a combination of different space physical parameters and the auroral oval boundary.

Responses of the magnetic latitudes of the poleward

The IMF can affect the auroral oval boundary through different space processes. In this experiment, the responses of different IMF components to the auroral oval boundary are shown in Fig. 5. The different colour and shape markers represent different MLT sectors. The vertical error bars represent one-eighth of the standard deviation from the mean value of the auroral oval boundary position, and the horizontal error bars represent the standard deviation from the mean value of the different IMF components in all of the binned data. Therefore, the length of the vertical error bar is fixed and the length of the horizontal error bar is changeable because of the different standard deviation in all of the binned data.

From Fig. 5, we can see that the poleward and equatorward boundaries in each
MLT sector show a step-by-step poleward displacement with the increase in the
IMF

Responses of the magnetic latitudes of the poleward

For the sake of finding the variation trend of the auroral oval boundary with the
change in solar wind parameters, including solar wind density (

From Fig. 6, both the poleward and equatorward boundaries shrink at
21:00–06:00 MLT when the value of

Response of the magnetic latitude of poleward

In this experiment, the average tendency of the poleward and equatorward boundaries influenced by geomagnetic indexes (AE, AL, AU) is shown in Fig. 7.

As we can see from Fig. 7, in every MLT sector, the poleward and equatorward boundaries move to low magnetic latitude with the ascending AE and AU indexes, while the poleward and equatorward boundaries extend to high magnetic latitude with the ascent of the AL index. The AE index is often used to characterize the strength of substorm activity in the magnetosphere. Therefore, it can be considered that the auroral oval extends to the Equator due to the enhanced substorm activity. Furthermore, the amount of energy enters the magnetotail along with the strengthening of substorm activity. This means that the AE index will increase when energy in the magnetotail is released through a substorm, which is coincident with our findings about AE from Fig. 7.

The MAE influenced by different space physical parameters. The bigger MAE values in both the poleward and equatorward boundaries were written in bold font compared with others. The values written in bold font representing its corresponding space physical parameters had more influence on poleward and equatorward boundaries.

As we know, most of the studies on how the space physical parameters affect the auroral oval boundary focus on solar wind parameters, geomagnetic indexes, and IMF components. There have been lots of corresponding conclusions about the influence of those space physical parameters on the auroral oval boundary up to now. Nonetheless, how the other space physical parameters not mentioned above affect the auroral oval location has not been addressed. In order to further explore the variation of the auroral oval boundary influenced by different space physical parameters, experiment 4 is performed. In experiment 4, we send one physical parameter selected from Table 1 at the present moment and the coordinates of auroral oval boundary points at the previous moment to our prediction model, and the outputs of our model are 48 coordinate values of auroral oval boundary points and the MAE between real boundaries and predicted boundaries. The MAE values of poleward and equatorward boundaries influenced by different space physical parameters are given in Table 3. We can infer the response of the auroral oval boundary to 18 space physical parameters through the different MAE values of these space physical parameters.

The MAE values of the boundary positions are 1.6076 and 1.4545 respectively when we only
use boundary positions at the previous moment to predict poleward and
equatorward boundaries. We take this MAE as standard, called S-MAE. Compared
with the S-MAE, we can see that the MAE increases about 1.9 % for the poleward
boundary by adding any one space physical parameter to the input of our model
from Table 3. Meanwhile, the value of MAE for the equatorward boundary is
between

The Pearson correlation coefficient of all 18 space physical parameters from December 1996 to March 1997.

In order to analyse the influence of space physical parameters on the auroral
oval efficiently, we not only consider the effect of each space physical
parameter on the auroral oval boundary, but also take the effect on the auroral oval
boundary with different combinations of space physical parameters into
account in experiment 5. As a result, we first calculate the correlations of
all 18 space physical parameters using the Pearson correlation coefficient,
which is a statistical value that reflects the degree of linear correlation
between two variables. The Pearson correlation coefficient of two variables
(

We can see that AE has a strong positive correlation with AU and PC and that AL
has a strong negative correlation with AE, AU, and PC from Table 4. The linear
correlation coefficient between AE and AL is

The MAE influenced by different combinations of space physical parameters.

Response of the magnetic latitude of poleward

The MAE influenced by different combinations of space physical parameters.

According to Table 4, there was an obvious correlation among the following
space physical parameters. For the solar wind parameters,

As a summary, it can be seen that these space physical parameters, which
include

In this paper, we establish a model to measure the relationship between
space physical parameters from the OMNI dataset on the NASA website and poleward and
equatorward auroral oval boundaries based on a deep learning network. Our
model overcomes some drawbacks in this field, such as some prediction
methods based on statistics and a few space physical parameters. Those
methods are not very suitable for the complex and changeable space physical
data. For our model, the inputs are 18 space physical parameters and the 48 coordinates value of auroral oval boundary points at the previous moment, and
we can obtain positions of poleward and equatorward boundaries at 24 MLTs
from our well-trained model. At last, our experiment results show that the
model proposed in this paper can better reflect the relationship between
space physical parameters and an auroral oval boundary. Therefore, it should be
useful to predict the position of an auroral oval boundary. In addition, we
analyse the effect of all 18 space physical parameters on the location of
an auroral oval boundary based on several statistical and prediction
experiments. It can be shown that different parameters have different effects
on the auroral oval boundary from our experiments. Some space physical
parameters,

All original aurora images and space physical parameters data were downloaded from the NASA (

The supplement related to this article is available online at:

The methodology has been developed by all of the authors. HY have coded and run the experiments. All the authors have discussed the theory and the interpretation of the results and edited the manuscript.

The authors declare that they have no conflict of interest.

The authors thank Jianjun Liu for his comments and suggestions.

This research has been supported by the National Natural Science Foundation of China (grant nos. 61572384, 41874195, 41831072, 41504116, and 41674169), the Shaanxi Key Technologies Research Program (grant no. 2017KW-017), China's Postdoctoral Fund First-class Funding (grant no. 2014M560752), the Shaanxi Province Postdoctoral Science Fund (grant no. JBG150225), and the National Key Research and Development Program of China (grant nos. 2016QY01W0200 and 2018YFC1407300).

This paper was edited by Jörg Büchner and reviewed by Unnikrishnan Kaleekkal and one anonymous referee.