An end-to-end method of Coronal Mass Ejections detection

Abstract

Coronal Mass Ejections (CMEs) impact heavily on coronal activity, space weather and many interplanetary disturbance, so the detection of CMEs are important for space weather disaster prevention and reduction. The traditional methods use man-made features or predefined threshold to solve this problem. Despite the great progress in the detection of CMEs, it is still a challenging problem due to the following three aspects: Firstly, the early and late stage of the CMEs phenomenon is very weak, and the traditional image processing based method can not detect this weak CMEs well. Secondly, the noises from comet, planets and other stars can affect the detection of CMEs. Thirdly, the CMEs are complex and amorphous, and they are different in shapes, textures, grayscales, scales and so on. Because of these difficulties, it is difficult to detect CMEs well by the traditional image processing method without modeling the CMEs. With the development of convolutional neural networks (CNNs), it is possible to develop deep neural networks based CMEs detection models to better solve this problem. For realizing this, this paper presents an end-to-end detection method of Coronal Mass Ejections detection: We design a deep neural network with 4 convolution layers, 1 full connection layer and 1 output layer. This deep neural network can automatically extract the image features that are suitable to describe the Coronal Mass Ejections, and can establish the CMEs detection model based on the extracted features. In order to achieve good performance, we construct two datasets, one is mainly made up of strong CMEs, and the other is made up of weak or dark CMEs. Training is first done on the strong CMEs to obtain the initial CMEs detection model. Based on the initial model established on the strong CMEs, finetuning is used on the weak CMEs to acquire the final CMEs detection model. By using this scheme, training efficiency and good performance can be guaranteed. In addition, the process is able to achieve selection of features and setting of classification rules, which can realize conveniently the end-to-end detection from data to results. The experimental results show that this method can effectively detect CMEs. The accuracy of our method is 100% on the strong CME dataset, and 91.54% on the weak CME dataset. The overall accuracy on our dataset is 98.05%. Finally, we test our method with the real data on May 2007 using LASCO catalog as the groudtruth. Experimental results show that our method achieves state-of-the art performance comparing with the usually used catalog, SEEDS, CACTUS. The significance of our work is two-fold. On one hand, our method is end-to-end method which can select optimal features for the detection task. According to the theory of deep neural networks, the selected features are usually those that the human vision uses. Therefore, the obtained features can depicts CMEs well and can be used by other methods. On the other hand, the good performance of our method proves the depiction ability of the CNN for CME. So it is natural that CNN based description for the CME can also have good performance for the study of the modeling of the evolution of the CMEs.