Detection And Tracking Of Solar Active Regions Based On YOLOv3 And DeepSort

Active regions(ARs)are the primary energy sources of various solar activities.The violent solar activities adversely affect human living environment.Therefore,accurate detection and tracking of ARs are very important for monitoring and forecasting the space weather.The tracking of ARs is a typical multi-object tracking problem.According to the task requirements of ARs' tracking and the characteristics of detection-based tracking(DBT)mode,we selected the DBT mode.Firstly,the detection algorithm is used to detect the ARs on the solar full-disk magnetograms,and then the tracking algorithm is used to track the detection results.The traditional image processing technologies have been mainly used to detect and track the solar ARs.There are some problems in detection and tracking ARs using these methods:(1)the bipolar of one AR is mis-detected as multiple ARs;(2)the multiple ARs that are closer to each other are mis-detected as an AR;(3)it is failure to capture the new ARs and terminate the disappeared ARs in time.In this thesis,we propose an AR detection and tracking method(ARDTM)based on the deep learning model comprising of the YOLOv3-spp and DeepSort.The main works are as follows:(1)The ARs detection dataset is established;(2)The detection method of the ARs based on YOLOv3 is studied.YOLOv3-spp is adopted as the detection method of ARs.This method adds Spatial Pyramid Pooling on the basis of YOLOv3,which effectively improves the detection performance and accuracy;(3)The ARs tracking dataset is established;(4)The ARs tracking method based on the latitudinal differential rotation law and DeepSort is studied.This tracking method has improved the recognition network,object prediction,object matching and termination conditions in DeepSort.These improvements effectively improve the tracking performance of the algorithm in ARs.The ARDTM method solves the problem that one AR is mis-detected as multiple ARs,or multiple ARs are mis-detected as an AR.Besides that,it captures the new ARs and terminates the disappeared ARs in time.The method improves the precision and recall of detecting and tracking ARs.The training dataset of this thesis only uses the HMI data,but it can detect and track the ARs well whether HMI data or MDI dataare tested.It shows good generalization.On the ARs detection dataset,the recall,the precision and the AP of this method reach 92.77%,91.70%,and 91.04%,respectively.In the course of tracking the ARs,the multi-object detection accuracy(MOTA)of sequence data at 12-minute intervals reaches 91.2%.The MOTA of the 24-hour interval sequence data reaches 82.7%.The MOTA on the two sequence data reach90.2%.The results show that the ARDTM method can be used for detecting and tracking ARs in the solar full-disk magnetograms observed from different telescopes,or different time interval series-images.In order to further improve the tracking performance of short interval sequence data,this thesis improves the ARDTM for short interval(12-36 minutes)sequence data by adding local weighted linear regression(LWLR).This method is called LWLR-ARDTM.The result shows that it is suitable for tracking the ARs in sequence data with a time interval of 12-36 minutes.On the sequence data with 12 minute intervals,MOTA reaches 92.7%.It is higher 1.5% than that of ARDTM.