An algorithm to derive CME 3D parametersbased on machine learning and double view-point observation

Coronal Mass Ejections (CMEs) are the major sources of severe space weather events, causing potential damages to orbital and ground assets including satellites, space stations and power grids. To avoid the huge economic losses, it is crucial to understand the propagation of CMEs and derive physical parameters especially in 3-dimension for better prediction of CME propagation. We have developed an algorithm to automatically reconstruct CME structure based on double view-point observations and machine learning technique. The algorithm consists of three steps: region acquirement, model construction, function optimization. First, we train two Convolutional Neural Networks (CNNs) to identify the CME in visual observations from the Large Angle Spectrometric Coronagraph (LASCO) onboard the Solar and Heliospheric Observatory (SOHO) and the Sun Earth Connection Coronal and Heliospheric Investigation (SECCHI) COR-2 Coronagraph onboard the Solar TErrestrial RElations Observatory (STEREO), respectively. The CME region information is then leveraged with the Principal Component Analysis algorithm and Otsu's method. Next, we establish the Graduated Cylindrical Shell (GCS) model and project it into the field of view of the coronagraphs. In the final step, we construct a function to measure the difference between the image of the GCS model and the CME region. Then the optimal 3D CME parameters can be obtained. Several CME events are chosen to show the accuracy and effectiveness of our method. We also conduct a statistical analysis on 127 CME events from 2007 to 2014 to investigate the 2D and 3D parameters of CMEs. Our method can be used to provide CME initial parameters in magnetohydrodymic simulations for accurate prediction and understanding of CME.