A Deep Learning-based Coronal Mass Ejection Detection Model

Coronal Mass Ejection(CME)is an intense space weather phenomenon in which Coronal Mass Ejection from the sun's corona is emitted into interplanetary space.CME has a serious impact on space weather and human life.So it is of great significance to improve the detection effect of CME to forecast CME and ensure the safety of human production and life.Although many methods for detection of CME have been explored by scholars,the existed methods mostly use artificial defined features and artificial defined thresholds to detect CME.Due to the fact that artificially defined features cannot represent CME well and it is difficult to select a universal threshold,the detection effect of existed CME detection methods on CME needs to be improved.Target detection and semantic segmentation methods based on deep learning have achieved good results in the detection and segmentation of natural image targets.For example,Faster R-CNN has a high detection accuracy rate on the PASCAL VOC and COCO data sets,and the detection results achieved are better than traditional target detection methods and previous deep learning methods.As a semantic segmentation method of deep learning,U-Net has achieved good results in target segmentation of medical images.This paper proposes a CME detection method based on deep learning based on the advantages of Faster R-CNN and U-Net.Faster R-CNN was used to detect the general region of CME,and U-Net was used to segment the exact location of CME.The CME detection process of this method: 1.Use Faster R-CNN to perform target detection on the input corona image to obtain a rectangular area containing CME targets;2.Use U-Net network to process the rectangular area output by Faster R-CNN to obtain CME target at pixel level.3.The CME catalog generation module generates the CME catalog of this article according to the target frame obtained by the target detection and the segmented area obtained by the semantic segmentation network.There are the following differences between the corona image and the natural image: 1.Due to the accuracy of the corona camera,the corona image contains less information such as color and texture;2.The CME and other bright noises in the corona image(such as planets,satellites,and random noise)Etc.)The degree of discrimination is not high;3.The corona image series has time continuity,and the image similarity of two adjacent frames is relatively high.Based on the characteristics of the corona image,this paper optimizes the Faster R-CNN network for target detection and the U-Net network for semantic segmentation.In order to utilize the shallow information and temporal continuity of the corona image,this paper uses Res Net as a feature extraction network and fuses the feature maps of three consecutive frames of images in Faster R-CNN.In order to match the size of the CME target,this paper uses the K-means algorithm to calculate the size information of the CME target,and then sets the anchor parameters of Faster R-CNN according to the size information.In order to combine the deep and shallow information of the corona image,we add a residual network to the encoder of the U-Net network.In order to increase the utilization weight of useful information,this paper adds an attention mechanism SE module between the encoder and decoder of the U-Net network.Since there is currently no data set for CME target detection and semantic segmentation,this paper first combines three famous CME catalog information such as the Coordinated Data Analysis Workshop Data Center(CDAW),the Solar Eruptive Event Detection System(SEEDS)and the Compute Aided CME Tracking software package(CACTus)to manually annotate the target detection data set containing 9113 corona images.Then cut out the CME target area from the corona image according to the target detection frame,and use the labelme tool to CME mark the CME target area to make a semantic segmentation data set,and obtain a semantic segmentation data set of 120 images.The target detection data set is sent to the Faster R-CNN network for training,and the semantic segmentation data set is sent to the U-Net network for training,so as to obtain a target detection and semantic segmentation model suitable for CME detection.In order to verify the CME detection effect of the method in this paper,the detection results of the method in this paper were compared with the SEEDS and Cactus directories using the manually labeled CDAW directory as the reference.Using the CME labeled data in June 2007 as the test set,the algorithm in this paper detected 22 out of 22 strong CME events and 138 out of 151 weak CME events,and the detection errors of characteristic parameters such as center Angle and Angle width of CME events were within 5 degrees and 10 degrees,respectively.Through comparison,the proposed method has the following advantages: 1.The detection accuracy of the proposed algorithm for strong CME and weak CME reaches 95.5%and 94.7%,respectively.The detection accuracy of all CMEs was 94.8%,which was6.4%,77.4% and 81.9% higher than that of Seeds and Cactus.2.can detect the CDAW directory missing weak CME event;3.The parameters of CME center Angle and Angle width detected by this paper are more accurate than those of the CACTus and SEEDS catalogs.