Semi-Supervised-Based Dayside Aurora Image Classification

The aurora phenomenon is formed by solar wind which collides with the atoms in the upper atmosphere over the earth, and it is the only visible phenomenon with the geophysical characteristics by the naked eyes in the Polar Regions. The observation of morphology and evolution of the aurora can get a wealth of information about the magnetosphere and the Sun-Earth magnetic actions which is helpful for us to study the ways and the extent of solar activity on the Earth, and have a great significance to understand how the process changes the space weather.Pattern recognition is becoming a new method of aurora image classification and still at the early stage. There are a lot of problems need to be solved, especially not taking full advantage of the unlabeled samples.In the pattern recognition, supervised learning requires a lot of labeled samples which spend huge human labor to label them and the labels are highly subjective. On the other hand, there are not accurate models for unsupervised clustering. For these problems, this paper uses a method based on semi-supervised learning: such as, semi-supervised expectative maximum, self-training and graph based semi-supervised algorithm. Because the forms of aurora images are very complex, including the shape of twisted arc-shaped, ray-like, block, and curtain-like, we use the local binary pattern to extract the features of the aurora images.At last, this paper uses three semi-supervised learning algorithms to classify the static aurora images. The experimental results show that in the case of only a few labeled samples, use a lot of unlabeled samples to classify the all-sky aurora images can get a satisfactory result.